Although a significant number of security techniques are currently under development to counter complex video manipulation methods and security attacks, despite these efforts, a notable gap persists in understanding the evolving landscape of security techniques for video data. This paper aims to bridge that gap by offering an extensive review of video data security, focusing on attack analysis and eleven distinct security approaches while considering recent technological advancements. While existing reviews tend to be outdated [11,12], fail to incorporate recent developments, or focus primarily on specific use cases [13–16]. This article is crafted to be understandable by any reader, irrespective of their technical expertise. It aims to consolidate existing knowledge, provide insights, and guide the development of new solutions in the field of video data security. The findings of this survey will provide reference to researchers seeking in-depth insights into security strategies, as well as key challenges and future trends unique to video data security in their ongoing research.

The primary contributions of this paper are outlined as follows:

Analysis of various manipulation techniques that can be used to alter video footage. Additionally, we examine attacks targeting security services such as integrity, authenticity, and confidentiality of video content, as well as disruptions to video delivery services.

The survey structure, as illustrated in Fig. 1, comprises the following sections: Section 2 discusses the background of the study and the associated challenges in securing video data. Section 3 explores how videos can be manipulated, followed by an explanation of several attacks on videos and their impacts in Section 4. Section 5, divided into twelve subsections, examines different security techniques for safeguarding videos. The pros and cons are discussed in Section 6, followed by future trends in Section 7. Finally, Section 8 concludes the paper.

Confidentiality is a critical security principle, particularly in the context of video, as it primarily ensures the secrecy of data. Below are the attacks that can compromise the confidentiality of video data.

Interception Attack: In an interception attack, an unauthorised individual attempts to gain access to valuable information transmitted between two parties [28]. The hacker may intercept sensitive data by eavesdropping on the transmission channel, using specialised software to capture data, physically tapping into network cables, or analysing network traffic. In [29], the authors proposed the PriSE method, which restricts access only to authorized users, thereby ensuring video protection. Furthermore, vulnerability assessments aid in detecting and preventing interception attacks.

Eavesdropping Attack: In such an attack, hackers tap into the network cables to eavesdrop on confidential data transmitted over private communication channels without the knowledge or consent of the parties involved [30]. To address this, the authors [31] proposed a hybrid encryption technique for securing video conferencing. Combining classical networks and quantum keys helps protect the transmitted data, making it difficult for hackers to decrypt the information even if intercepted. The encrypted scheme can also efficiently resist high-performance computational threats such as eavesdropping.

Hijacking Attack: The hijacking attack occurs when a hacker tricks a user into clicking on a hidden or disguised link or button, often leading to unintended actions such as downloading malware or making fraudulent purchases [32]. There are several types of hijacking attacks, such as session hijacking, where the attacker steals the user’s session ID (session identifier) or cookies, and DNS (Domain Name System) hijacking, where the DNS settings are modified to redirect traffic to a malicious website. In [33], a radio frequency identification technique is presented to prevent unauthorised access.

Man-in-the-Middle Attack: In a man-in-the-middle (MITM) attack, the hacker positions themselves between two parties’ transmissions and can easily read messages that are being exchanged [34]. The middleman can inject or modify new messages into the conversation to impersonate one of the parties. To reduce the MITM attack, an enhanced version of the Discrete-Logarithm Problem (DLP) was proposed to secure the communication channel [35]. The attacker cannot manipulate or impersonate the exchanged information, nor can they obtain the encryption key. Fig. 5 illustrates the implementation of an MITM attack.

Interference Attack: In an interference attack [36], the hacker disrupts the normal functioning of communication by injecting noise or other unwanted signals. A radio frequency jammer can be used to transmit signals at the same frequency as the communication channel, disrupting the transmission. These attacks pose a significant problem in wireless communication, where communication signals are vulnerable to interference from other sources.

Integrity ensures that video data remains accurate, consistent, authentic, and reliable in video processing. Below are the attacks that can compromise the trustworthiness and reliability of video content.

Replays Attack: In a replay attack [37], the attacker captures a valid data transmission, such as a login request or a video transaction, then resends it to the server later. If the server accepts the replayed data transmission, the attacker can conduct malicious actions such as accessing sensitive information or transferring illegal video data. To prevent replay attacks, a modified high-frequency descriptor technique was used for video streams [38], thus ensuring the integrity of the data being transmitted.

Packet Injection Attack: Packet injection attacks can be executed through various methods, including the use of network sniffers, physical wiretaps, or malware [39]. In this attack, hackers send packets that mimic legitimate ones to gain illegal access to a network or compromise a device’s security. The hacker can modify or delete frames from a video sequence. In [40], the authors discussed encryption schemes that can protect data from interception, make it challenging for attackers to modify information, and restrict access only to authorised users.

Fabrication Attack: The fabrication attack involves injecting false data, such as fake video credentials, to manipulate the originality of a video sequence. The attack can be carried out in various ways, such as using social engineering tactics to trick users into providing sensitive information or exploiting vulnerabilities in software or hardware. To prevent fabrication attacks, the paper utilised YOLO (You Only Look Once) and SSD (Single Shot multibox Detector) algorithms [41]. The proposed method was effective against fabrication attacks and ensured the integrity of the video data.

Message Deletion Attack: A message deletion attack may involve the manipulation or deletion of video frames or metadata within a video stream [42]. A hacker might delete metadata within a video file, such as time codes or camera information, to conceal evidence of malicious activity or hinder the viewer’s ability to interpret the video accurately. Therefore, an efficient detection scheme was introduced by [43] that can prevent message deletion attacks.

Payload Attack: These attacks target the integrity of the video system, as discussed in [44]. Hackers can embed unwanted scripts or executable code within video files. When users open or play the injected video, the malware payload is executed, resulting in the installation of malicious software, data breaches, and system compromise. Hackers may leverage vulnerabilities to deliver payloads that exploit weaknesses and gain unauthorised access, compromising user privacy. The following measures can be implemented to mitigate payload attacks: patch management, secure coding practices, anti-malware protection, user awareness and education, sandboxing and virtualisation, and content verification.

Masquerade Attack: The hacker embedded the masquerade attack to impersonate or deceive users as trusted video sources by spoofing their identities [45]. They may forge video stream metadata like addresses, timestamps, source IP (Internet Protocol), and digital signatures to trick the video processing system into processing and accepting the video from an unauthorised source. This attack can be secured through secure authentication, access control, digital signatures, encryption and secure protocols, user training, content integrity verification, system monitoring, and anomaly detection.

Availability ensures that videos are available for authorised users and can compromise in the following ways:

Denial-of-Service Attack: A denial-of-service (DoS) attack targeted at a video sequence can result in significant disruption to video streaming [46]. There are several ways in which a DoS attack can impact video processing: (1) Network congestion: flooding a large number of requests to the network can cause congestion, making the video unwatchable; (2) Resource depletion: consuming resources like CPU (Central Processing Unit) and memory can cause the system to crash, rendering the video unresponsive; (3) Packet fragmentation attack: sending a huge number of fragmented packets to the video system can cause the system to become unstable.

Jamming Attack: A jamming attack refers to the intentional disruption of wireless video transmission by transmitting radio signals at the same frequency as the video signal [47]. Through this attack, video streams may be disrupted, resulting in pixelation, delays, prevention of monitoring critical events or areas, complete loss of signal, and loss of frames in live video streaming. Ashourian et al. proposed the Karhunen–Loeve transform (KLT) technique, based on the spiked matrix model and wigner matrix, which can help minimise the impact of jamming attacks, as discussed in [48].

Routing and Command and Control (C&C) Attack: In a routing and C&C attack, an attacker can manipulate the destination route, reroute video traffic to a malicious destination, and gain control of the video system [49]. This allows them to execute commands to access, manipulate, or steal video data, potentially obtaining sensitive information or using it for blackmail.

It refers to the illegal attempt to gain access to a service or systems by exploiting vulnerabilities using the following methods.

GCS (Ground Control Station) Spoofing Attack: In a GCS spoofing attack, a hacker impersonates a Ground Control Station in a video-based unmanned aerial vehicle (UAV) system. The GCS controls the UAV and receives video data from it. In such an attack, the hacker sends false signals to the UAV, causing it to follow commands from the hacker instead of the legitimate GCS. Thus, the hacker can take control of the UAV and gain access to the video data being transmitted from the UAV’s camera. Wei et al. presented a semantic-based method that can detect the spoofing attack from the UAV [50].

Machine learning attacks refer to utilising vulnerabilities in the system using various ML algorithms and techniques, as discussed below.

Spam Filtering Attack: Evasion attacks, or spam filtering attacks, are not typically delivered directly to the user’s inbox [51]. Hackers upload videos containing offensive or harmful content. They spam a large volume of data to make it difficult for users to find legitimate content, as discussed in [52]. These videos may be created using editing techniques or deepfake technology to appear trustworthy. Therefore, these attacks on video platforms and streaming services can be handled through collaborative filtering, content moderation, video analysis and classification, user reporting and feedback, and enhanced video metadata.

Intrusion Detection Attack: An intrusion attack involves uploading and streaming videos that contain malicious material or unauthorised content [53]. In such an attack, the hacker can modify or tamper with the video content by altering the video frames and metadata, deceiving the video processing system, and manipulating the recorded videos. The attack can be detected by system monitoring, video authentication, real-time video stream analysis, access control and encryption, and metadata analysis.

Fraud Detection Attack: This attack can be employed for fraudulent purposes by creating realistic videos that mislead users or manipulate information [20]. Hackers may create or distribute videos that promote scams, illegal activities, or counterfeit products. To combat these attacks, it is necessary to update security measures regularly and stay informed about emerging fraud techniques. Measures such as real-time monitoring, video authenticity verification, facial recognition and biometrics, user feedback and reporting, and behavior analysis can help in this regard.

Network Traffic Detection Attack: In this particular attack, hackers can intercept video traffic between the source and destination or modify the video content [54]. They can use packet sniffing tools to analyse traffic patterns associated with video systems to gain insights into system vulnerabilities or sensitive information. Therefore, these attacks on video platforms and streaming services can be handled through encryption and secure protocols, traffic monitoring and analysis, access control and authentication, intrusion detection/prevention (ID/IP), and network segmentation.

Table 1 highlights some of the attacks, their impact, and detection techniques while details are provided above.

To safeguard video data, researchers have developed various methods that have been utilised for decades, and yet they continue to evolve to achieve data protection and privacy preservation. These techniques include blurring, pixelisation, morphing, encryption, and others, as described in Table 2, primarily focused on offline videos. However, with significant growth in live streaming, communities have discussions about implementing similar for live video. The paper [71], introduced a privacy protection method named Face Pixelation in Video Live Streaming (FPVLS) to safeguard personal privacy rights and foster the healthy growth of the video live streaming industry. The related work also demonstrates that some of these approaches do not pay attention to intelligibility and reversibility, memory-constrained hardware architectures, and low network bandwidths, not to mention privacy aspects. The weaknesses of some methods other than encryption, such as pixelation and blurring, are highlighted in [72].

Nowadays, when media manipulation and deepfake are prevalent, encryption is vital for safeguarding video content’s integrity and authenticity. Cryptographic algorithms such as Data Encryption Standard (DES), Advanced Encryption Standard (AES), Rivest-Shamir-Adleman (RSA), etc., have been used to provide security since 1970. To mitigate the current challenges, an improved version of AES was introduced by [80], featuring a mix-row operation replacing sub-byte and shift-row operations to reduce runtime. While this scheme fully encrypts input video, its encryption time complexity is notably higher. A technique for securing video content and verifying receiver identity during online video streaming was proposed using a hybrid cryptography approach integrating AES, RSA, and Elliptic-curve cryptography (ECC) algorithms [81]. This hybrid model, combined with dynamically generated keys, significantly enhances security.

Blowfish, known for its efficiency in symmetric cipher algorithms, was employed by utilising key lengths ranging from 32 to 448 bits [82]. However, its susceptibility to weak key problems introduced complexity. To address this, the Blowfish algorithm was modified by altering the F-function structure, reducing the number of S-boxes to two, which enhanced efficiency and lowered computational costs [83]. In another study [84], the authors utilised the Blowfish algorithm to encrypt sensitive information, concealing it within steganographic images using the Least Significant Bit (LSB) technique. In [85], employed AES, the Vernam encryption algorithm, and LSB schemes to convert information into encrypted, incomprehensible forms. To solve the problem of key distribution in shared secret keys, a secure system based on cryptography and authentication was introduced by utilising elliptic-curve Diffie-Hellman (ECDH) [86]. The authors explored three systems aimed at securing the integration of encryption and authentication to ensure the authenticity and confidentiality of information exchanged between IoT (Internet of Things) devices over untrusted communication channels [86].

Chaotic techniques leverage mathematical functions to generate randomness or scramble data, providing promising solutions for video security. However, their effectiveness is constrained by limitations arising from their inherently less dynamic behavior. Nevertheless, they have been employed in video security with varying success. An intertwining logistic map scheme based on the cosine transformation of video was introduced by [87]. Permutation is applied to each frame to reduce pixel correlation; subsequently, the frames are rotated 90 degrees counterclockwise. Finally, a random substitution scheme is applied to make changes column-wise and row-wise, with frames bounced according to a frame selection key. In contrast, the authors [88] utilised improved chaotic maps and parallel compressive sensing techniques to efficiently encrypt/decrypt videos without compromising quality. A logistic Tent Infinite Collapse Map (LT-ICM) provides unpredictability, sensitivity, complex behavior, and a broader chaotic range for controlling parameters. Additionally, they proposed a substitution method consisting of exclusive-OR (XOR), modular addition operations, and circular shifts to substitute frames using a two-dimensional LT-ICM sequence. Recently, Hadjadj et al. [89] introduced a chaotic-based approach that relies on a unified Lorenz chaotic generator and the One-Time Pad (OTP) technique, altering the behavior of chaotic systems. In case of an attack, this technique provides dynamic reconfiguration and resilience against adversaries.

In [90], a data compression and mapping algorithm was introduced to implement data encryption via video processing. Initially, moving objects and backgrounds were extracted through a video composition model. Then, a data encryption model was applied to facilitate effective data transmission, safeguarding against the extraction of background and moving target information. In another endeavor to secure video communication over network environments, Lin et al. [91] presented a synchronisation controller scheme. This scheme facilitates the transmission of both slave and master chaotic systems, enabling dynamic chaotic and synchronised random number generation simultaneously. While this approach efficiently secures video communication, it does not fully address robustness and reversibility concerns. To overcome these challenges, a multi-domain characteristic and Logistic-chaotic scrambling have been utilised to embed elements in encrypted videos, denoted as element-I [92]. The videos were further encoded using Bose–Chaudhuri–Hocquenghem code. Subsequently, element-I, containing robustly embedded label information, modified the amplitude of element-I, enabling the recovery of lossless auxiliary information represented as element-II. The auxiliary information was reversely embedded into element-II using traditional histogram shifting.

In another work [93], a joint operation of two linear chaotic-symmetric maps and one chaotic-tent map has been proposed. This technique permuted each frame’s pixel positions and shuffled linear symmetric-chaotic sequences using a P-box. More recently, Es-Sabry et al. [94] introduced an enhanced method for encrypting 32-bit color images using four 1D (one-dimensional) chaotic maps: Tent, Logistic, Sine, and Chebyshev. These maps carefully filled the four vectors in a cryptographic scheme, assigning unique integers between 0 and 255. They further enhance security by securing each pixel’s digits using the Four-square encryption. Additionally, they introduce an additional layer of complexity during the obfuscation stage by intentionally rearranging all pixel positions using the Arnold Cat Map transformation. An encryption system based on cellular automata (CA) and S-boxes has been proposed [95], which integrates a four-dimensional (4D) memristive hyperchaos with an additional exceptional chaotic variety, thereby enhancing uncertainty and ergodicity.

With the proliferation of the IoT across diverse domains like smart homes, cities, industries, and healthcare, conventional security techniques face inefficiencies when applied to IoT-enabled devices. These challenges arise from the computational intensity, bitrate overhead, and bandwidth utilisation inherent in conventional approaches. As a remedy, lightweight and efficient security solutions become imperative. This section delves into proposed lightweight ciphers and security techniques specifically designed to safeguard video data in the Internet of Multimedia Things (IoMT).

Lightweight Cryptographic Approaches: Recently, researchers and standardisation bodies have shown much interest in designing lightweight algorithms for secure end-to-end communication. Thus, algorithms have been proposed that employ existing SOTA encryption algorithms by reducing the number of rounds and new lightweight cryptographic algorithms. In [96], a dynamic key-dependent LoRCA cipher that comprised two rounds, one for block and another for stream cipher, was presented. Each round was implemented for a single iteration to minimise resource and computational costs. LoRCA cipher demonstrated superior performance compared to Speck by 388%, Simon by 358%, and a 100% improvement over AES. A combined encryption technique that integrates the sigmoid logistic map, Kronecker XOR product, and Hill cipher algorithms was introduced in [97]. The algorithm initially shuffled the matrix rows to the right and then applied the Hill cipher to the resulting image. Subsequently, XOR operations were performed on each value of even or odd columns. Finally, the Kronecker XOR product and sigmoid logistic map were implemented on the resulting image to enhance security and resilience against various attacks. In another work, a SLEPX (Symmetric Cipher for Lightweight Encryption based on Permutation and EXlusive OR) algorithm was proposed to encrypt the Scalable High-Efficiency-Video Coding (HEVC) videos [98]. The authors claim that their scheme offers lower computational costs than XOR and provides security approximately equivalent to AES.

Selective Features-Based Techniques: To address complexity concerns, the study [99] proposed a Selective Encryption (SE) capable of encrypting moving objects using the AES algorithm. This technique achieves high-quality encrypted videos, analyses non-encrypted statistics, and minimises computational costs. In [100], a technique based on chaotic maps and singular-value decomposition (SVD) was proposed for the SE of video feeds. SVD was utilised to select significant parts of video frame feeds containing the most crucial information. The significant frames were encrypted using the PRESENT block cipher, generating an 80-bit key size using the logistic map. In [19], the authors presented a chaotic logistic map-based encryption scheme for encrypting human faces (selected as ROI). Initially, the face is identified using the YOLO v3 detection algorithm. Subsequently, the fast-block scrambling method is applied to scramble the ROI, followed by implementing a cipher to protect the facial representation. In [18], a selective bit-stream encryption based on the chaotic Arnold map (CAM) for scalable HEVC (SHVC) was introduced. This scheme encrypts/decrypts Sample Adaptive Offset, Motion Vector Difference (MVD), and Discrete Cosine Transform (DCT) bits in AES cipher feedback (AES-CFB) operation mode. The CAM-based SHVC-SE scheme minimised encoding time compared to AES-based SHVC-SE. In [101], a message-embedding encryption technique was introduced, allowing the extraction of message bits during video decryption. The embedding of messages is achieved by altering values in reference frames and motion vectors by swapping the values of x and y components, thereby scrambling the video result. Whereas, in [102], SE based on the Rivest Cipher 4 (RC4) was proposed, ensuring correct decryption even in cases of packet loss. The scheme shuffles the non-zero coefficients using a two-round-shifting algorithm, with Quantised Transform Coefficients (QTCs), intra-prediction modes (IPMs), and MVD chosen as selective parameters.

Table 3 provides a comprehensive summary of SOTA lightweight security techniques.

Security for IoMT: IoMT encompasses multimedia devices like cameras and sensors; however, hackers may access the IoMT network or individual devices, allowing them to intercept video streams, manipulate camera settings, or even disable cameras altogether. To address the issues, researchers proposed solutions to safeguard the videos captured by resource-constrained IoMT devices. In [67], the authors apply Fernet encryption to safeguard against unauthorised access and MITM attacks. Additionally, they utilise the Diffie-Hellman protocol for secure key exchange, while the SHA256 (Secure Hash Algorithm 256-bit) verifies the authenticity of video sequences. Another work [105] introduced a lightweight SE to secure videos within the IoMT infrastructure. The proposed SE method can selectively encrypt parameters such as the sign of MVD, UEGO (Unary code and Exp-Golomb code) suffixes, a sign of NZ-TC (non-zero transform coefficient), and dQP (difference of quantisation parameters). In [110], to tackle the space and time complexity constraints of IoT devices, multi-keyed logic and logical operations like AND, XOR, and OR are employed. Additionally, to ensure the security of videos during transmission over IoT-Fog, chaining logic, Op_codes (Operational codes), and HMAC (Hash-based Message Authentication Code) is utilised. In [111], the authors utilised the PRESENT and PRINCE cryptography algorithms within the default architecture of the Reon V-processor-core to establish robust security mechanisms for IoT applications. This approach aimed to achieve high performance and encryption results. Recently, the authors [112] proposed the Elman Neural-based Blowfish-Blockchain mechanism by integrating the features of Blowfish cryptography and the Elman Neural System. The approach involves two phases: firstly, the crypto analysis phase encrypts the dataset, and secondly, the monitoring phase removes attacks from the dataset. The encrypted datasets are then stored in a cloud server. The authors presented a modified DNA-based encryption algorithm tailored for IoT devices [113]. This scheme generates a purely random secret key, enabling efficient transposition, circular shifting operations, and substitution while managing high levels of information security. Modifying the DNA-based encryption scheme enhanced the data block and key space. More recently, a technique named RAFCA (Resource Allocation Functionality with Cluster Aggregation) for securing the transmission of surveillance videos was proposed, which is based on the permutation process of the video sequence over various mobile relays [114]. In contrast, another work on a 2D (two-dimensional) 4-scroll chaotic system that can easily manage the real-time movement of data is proposed [115].

In recent years, ML has garnered significant attention across various applications, particularly in cybersecurity, for detecting and predicting different types of attacks. However, leveraging its advanced capabilities also holds great potential in protecting video data. In [116], the authors applied unsupervised ML algorithm motion fusion to recognise the motion of objects, with foreground (FG) pixels and background (BG) pixels separated from each other using global thresholding. Encryption is then applied to these pixels with the ChaCha20 algorithm, where the key and nonce value are randomly generated. Another work [117] proposed an ENCVIDC ML technique that employed random forest (RF), K-nearest neighbors (K-NN), and support vector machine (SVM) algorithms to identify encrypted video with peak performance and accuracy of 98.13% in a second. Furthermore, as a subset of ML, Deep learning (DL) also benefits video security, incorporating enhanced accuracy, scalability, and automation. Therefore, researchers continually explore new DL-based methods to ensure video data security. For instance, Geng et al. [118] proposed an encryption algorithm utilising DL to detect target images from videos. This algorithm selects images based on important and non-important regions of interest, encrypting them to generate cipher images. In another study [119] the authors developed a deep convolutional neural networks (DCNN) technique by integrating the Bird-Swarm Algorithm (BSA) and Sine-Cosine Algorithm (SCA) to insert secret information into video frames. Keyframes are identified using Wavelet and Minkowski distances, with regions of interest detected by DCNN. This scheme is both efficient and less time-consuming. On the other hand, Kaczyński et al. [120] introduced watermarking into high-quality videos encoded with the H.265/HEVC codec, employing adjustable subsquares properties technique with deep neural networks (DNN). This approach yields superior results for high-quality videos compared to low-quality ones.

DNA-based encryption refers to a cryptographic approach in which cryptographic algorithms are designed to mimic the behavior of biological processes of DNA (Deoxyribo-Nucleic-Acid) molecules, such as encoding, decoding, replication, and mutation [121]. The idea behind DNA-based security is to leverage the complexity and robustness of data encryption, authentication, and access control. Researchers have explored DNA-based security techniques to assess their efficacy in securing video data. Reference [122] utilised the DNA sequence technique along with a 5D (five-dimensional) hyperchaotic system. The DNA sequence, represented by combinations of 0 and 1 s, is denoted as follows: adenine (A) represented as A = 00, cytosine (C) as C = 01, guanine (G) as G = 10, and thymine (T) as T = 11. These combinations are used to encrypt/decrypt a binary sequence. Instead, the study [123] introduced DNA sequences to generate intricate secrets and encrypted bits, thereby intensifying computation time and encryption capacity. In [124], the authors combined chaotic and DNA sequences. The scheme separated key and non-key video frames from a video sequence, with the DNA sequence employed to extract and encrypt keyframes. The authors [125] implemented DNA sequences in combination with Cellular Automata (CA) and chaotic systems on digital videos to address copyright and confidentiality concerns. The proposed scheme can be parallelly implemented on various CA rules due to one-dimensional chaotic automata based on DNA. The work [111] employed DNA cryptography and chaotic maps by generating three keys using a circle, tent, 3D (three-dimensional) logistic map, and Chebyshev in a multilevel fashion. These keys determine DNA’s encoding–decoding performance on the subblocks, used for row-column rotation of subblocks, and encryption is applied using DNA XOR operation [126] introduced a hybrid method by integrating the Mersenne Twister (MT), DNA, and Chaotic-Dynamical-Rossler (MT-DNA-Chaos), which provides enhanced security and improved efficiency of data randomness.

Drones, also known as unmanned aerial vehicles (UAVs), equipped with multispectral cameras and sensors, are utilised across various industries and applications, providing valuable insights, enhancing situational awareness [127], and facilitating live video feeds from locations and terrains that are challenging or impossible to access with traditional video-capturing methods [128]. However, they also pose potential security, safety, and privacy risks that must be addressed [129,130]. In [131], a communication of drone-based monitoring is secured through the AI-based Secure Communication and Classification for Drone-Enabled Emergency Monitoring Systems (AISCC-DE2MS). The proposed scheme uses the artificial gorilla troops optimiser (AGTO) algorithm to encrypt images using the ECC-based ElGamal encryption scheme. The AISCC-DE2MS can extract densely connected networks, hyperparameter tuning through penguin search optimisation (PESO), and perform long short-term memory (LSTM)-based classification, making it an effective tool for securing drone-based communication. The authors [132] introduce swarm-aided drone monitoring utilising distributed blockchain technology. In the proposed system, encrypted security protocols are utilised to certify security and ensure the confidentiality of information during transmission and the authenticity of UAV nodes’ identity. To mitigate attacks such as MITM and replay attacks, the authors proposed the ACPBS-IoT (Access Control Protocol for Battlefield Surveillance in drone-assisted Internet of Things) using AVISPA (Automated Validation of Internet Security Protocols and Applications) simulation tool and MIRACL (Multiprecision Integer and Rational Arithmetic Cryptographic Library) to calculate the execution time of Raspberry Pi 3 and the server [133]. In [134], control of electromagnetic fields (CEMF) technology is introduced to secure perimeters, addressing issues such as inflated alarm failure and vulnerability to weather conditions. The proposed system remains powered off until it detects people’s motion or another event. Autonomous drone technology provides much better intruder tracking and detection, overcoming false alarms and allowing cameras to be deactivated most of the time.

Homomorphic encryption (HE) is a cryptographic technique that enables operations to be performed directly on encrypted data, generating results that, when decrypted, are equivalent to those obtained from operations on plaintext [135]. This property makes HE ideal for securing video data, as it allows for secure processing, transmission, storage, and analytics of video data while maintaining its confidentiality and integrity, as presented in [136–138]. In [136], the authors proposed enhanced-fully homomorphic encryption (EFHE), utilising the fully homomorphic encryption (FHE) model. To authenticate videos in the cloud, an integrity-based FHE method was proposed by [139], where the video data was inaugurated to minimise time complexity, thus enhancing cloud services’ authentication and integrity. In [140], an architecture for enabling multimodal inference using homomorphic encryption more lightly and effectively was introduced. The Tensor Fusion Network-based HE technique was initially implemented to encrypt multimodal fusion features. After that, two methods for handling heterogeneous data, pre-expansion, and packaging, were employed, effectively reducing information traffic and temporal lag related to homomorphic computation. In [141], the authors described the Paillier encryption, based on semi-homomorphic encryption (SHE), which can measure encrypted data. While providing security guarantees and strong privacy, the proposed scheme required more energy and time. Priya et al. [142] also addressed video integrity, using the Paillier homomorphic cryptosystem for encrypting/decrypting original data. Although this scheme is secure, it increases computational complexity. In [143], a technique named RACE based on HE was proposed for encrypting/decrypting edge-to-edge communication. While providing sufficient security, it demanded more energy resources. In another study [144], the authors presented a HE scheme with pixel selection based on optimal metaheuristics (OMPS-HEVS). The introduced OMPS-HEVS converted frames and applied 2D-discrete wavelet decomposition to encode secret messages, increasing security and performance using the optimal homomorphic encryption (OHE) with the Jaya optimisation algorithm.

Watermarking is used to embed hidden information into digital media, typically a digital signal or pattern, to safeguard content integrity, authenticity, and ownership, along with providing copyright protection and content identification [145]. It can be applied at various video production and distribution stages, including content creation, post-production editing, and distribution. In videos, watermarking subtly inserts metadata such as copyright details or ownership rights into frames to identify the creator. Generally, three methods are used for watermarking video content [146]: (1) Spatial domain watermarking directly modifies the pixel values of the video frames to embed the watermark, typically used for visible watermarks overlaid on top of the video content [147]; (2) Frequency domain watermarking operates on frequency components like DCT coefficients, embedding watermarks in a less perceptible manner while maintaining robustness against signal processing operations [148]; (3) Spread Spectrum Watermarking spreads the watermark signal across the entire video signal using a pseudorandom sequence, with embedded watermarks extracted using a matching correlation process at the receiver [149]. In [150], the authors presented a hybrid compression-based digital watermarking technique that employs dual-tree complex-wavelet transform to decompose the video frame. The ECC encrypts and converts these encrypted images into binary bits and then inserts them at specified positions in the frame. The proposed technique minimises the size of the frame without affecting its quality. Another work [151] used a hybrid technique based on SE and watermarking. The scheme applies singular value decomposition and homomorphic transform in the discrete wavelet transform (DWT) to enhance the functionality of the watermarking method, achieving copyright protection and confidentiality of transmitted information. The study [152] implemented the block-based singular-value decomposition hybrid technique. The infrared and cornea frames are watermarked using SVD watermarking and dispatched through an erroneous wireless channel. The watermark scheme enhances wireless communication channels’ security, robustness, and detectability.

Cao et al. [153] presented a hyperchaotic Lorentz system-based watermarking algorithm to provide robustness and high imperceptibility. In [154], a combined watermarking and encryption based on the homomorphism scheme is proposed. The method can integrate both techniques into the same operand to provide higher security. However, this method is not helpful for Paillier encryption and Patchwork watermarking. The authors [155] proposed a tensor feature map technique based on a watermarking algorithm. The tensor feature map restrains the information of every frame, allowing watermarking to be distributed by inserting the watermark into the tensor feature map. The tensor feature map uses the discrete cosine and wavelet transforms for watermarking. Tian et al. [156] utilised a semi-fragile video watermarking technique based on chromatic-DCT. The scheme can increase bit rate, robustness, and video quality. The authors Zhang et al. [157] presented a watermarked-based time factor matrix adjusting the initial column of the time factor matrix and thus can protect against scaling, rotation, cropping, video compression, and frame deletion. The scheme converts the host video into three-factor matrices and a core tensor, representing the frame in the column, row, and time-axis direction by implementing Tucker decomposition. Du et al. [158] proposed a watermarking technique based on Tensor-QR decomposition (T-QR) and human-visual system (HVS). In [159], videos were protected by implementing a watermarking technique based on a pseudo-three-dimensional cosine transform (pseudo-3D-DCT), non-negative matrix factorisation (NMF), and non-subsampled contourlet transform (NSCT). The proposed technique provides invisibility of the watermark, robustness, and high ability against combined attacks.

Blockchain is an emerging technology that can store or manage a massive amount of data and prevent unauthorised access to the data. In [160], the authors proposed a two-level blockchain system to ensure the integrity of the videos. The proposed systems divide digital evidence into cold and hot blockchains. In the investigation process, the frequently changed information is stored in the hot blockchain, while unchanged information is stored in the cold blockchain. Another blockchain-based technique was proposed to distinguish the original videos from fake ones, thus ensuring authenticity [161]. To address issues related to centralised security systems, the authors [162] proposed a distributed-video surveillance method based on blockchain. The technique protects privacy, consumes fewer resources, maintains integrity, and manages blurring keys. The study [163] utilised blockchain based on the Merkle-Tree method, which can efficiently transmit video data, reduce required bandwidth, minimise storage costs, and safely synchronise CCTV (Closed-Circuit Television) video. In another work, the authors [164] proposed an algorithm based on the secure multiparty computation (SMC) blockchain that can accurately verify and authenticate video data token records. In [165], the authors provide design guidelines for blockchain-based DRM solutions that enable visible licensing of music frameworks, consistent and comprehensive rights metadata, and efficient and visible royalty distribution. Three methods are used to accomplish the solution: (1) putting rights metadata on a publicly distributed ledger, (2) using a consensus process on a blockchain with permission to validate metadata, and (3) using a smart contract to enforce royalty payouts using stablecoin. The study [166] authenticate the integrity of VVC (Versatile Video Coding) by implementing the concept of an ON and OFF chain. First, the video hash is calculated through HMAC; afterward, the calculated hash and the encryption key are stored on the ON chain. Second, the selective features of the video are encrypted through SLEPX and then stored on the OFF chain.

Digital Rights Management (DRM) refers to techniques for addressing intellectual property rights issues. It can prevent unauthorised copying, distribution, and modification of digital content [167]. Researchers have employed DRM for video data in a few ways beyond simply applying commercial DRM solutions. In [168], the authors proposed a SE for H.264/AVC videos integrated with DRM techniques. This method encrypts the video texture during compression using DCT coefficients, ensuring high encryption speed, robust security, and suitability for DRM and industrial applications. In [169], the authors introduced DRM by implementing Secure Digital Camera (SDC) rights at the sender end for video data. The scheme utilises digital watermarking and encryption to embed binary images into the video framework and authenticate integrity using a unique key. In [170], a DRM is presented to address copyright issues in video systems by employing the AES for encryption/decryption. In another study [171], a streaming-based DRM solution is proposed using the JSON Web Token (JWT) technique to verify individual identification. The server’s RSA secret key encrypts query variables, while the client provides a set of public and private keys. The video employs adaptive protection with AES-128, generating a new key with each performance to enhance security.

More recently, generative adversarial networks (GANs) offer a range of applications in video security, from detecting deepfake [70] for enhancing video quality, detecting anomalies [172], generating synthetic data [67], preserving privacy [173,174], and authenticating video content. By leveraging the capabilities of GANs, video security can become more robust, accurate, and reliable, addressing many challenges posed by modern security threats. For instance, in [70], the authors proposed a lightweight deepfake detection model for video conference applications to detect fake faces. It employs the MTCNN (Multi-task Cascaded Convolutional Networks) method for deepfake detection, demonstrating that the model can detect faces quickly and with acceptable accuracy, outperforming traditional methods. Additionally, the use of an Inception-Resnet model enhances performance results. In [67], the authors created synthetic data to comply with privacy regulations for objects/individuals and generate a large dataset that is difficult to collect in a real-time environment. In [174], the authors introduced a framework called Privacy-Protective-GAN (PP-GAN) for face de-identification that incorporates verification and regulator modules, addressing the limitations of traditional methods like the k-same framework, which suffers from low effectiveness and poor visual quality. The verification ensures that face recognition models cannot recognise the generated faces, while the regulator ensures that the generated faces retain the same structure as the input faces. This means that the output faces, while looking similar to the input, cannot be used to identify the original individual, thereby preserving privacy. Balancing these two aspects makes it a powerful tool for applications where privacy is a concern, such as in public surveillance or social media platforms. Whereas, in [175], a GAN-based technique is employed to train a synthetic dataset for deepfake videos, using 132,000 video frames extracted from John Oliver’s YouTube videos. In [176], the authors proposed a recurrent neural network-based technique to detect video manipulations, utilising 300 deepfake films from various video-hosting websites. An additional 300 videos randomly chosen from the HOHA dataset were also incorporated, resulting in a final dataset comprising 600 videos.

The use of AI and ML for video security is an actively growing research area, as discussed earlier. However, AI-driven detection, such as forgery detection, identifying real-time attacks, predictive threat modeling, and automatically responding to security incidents, remains complex and still lacks accuracy. Moreover, in cases where malicious actors manipulate or perturb input video data to deceive ML models, defenses against these adversarial attacks [204], particularly those resistant to multi-perturbation are still in their nascency.

Edge-based security refers to securing data by leveraging the edge-computing paradigm [205] to perform tasks such as encryption, decryption, content filtering, anonymisation, and threat intelligence directly at edge devices where the videos are generated. This ensures that sensitive video content remains secure throughout the video delivery process (End-to-end), thereby protecting individual privacy and ensuring regulatory compliance. For example, in [206], privacy concerns are addressed while enabling effective surveillance at the edge. However, edge devices typically have limited processing power and storage capacity compared to centralised servers. This can restrict the complexity and scale of security operations that can be performed at the edge due to computational and bandwidth requirements. These challenges necessitate the need for efficient, equally robust security measures and careful consideration of the specific requirements and constraints tailored to the edge computing paradigm while dealing with the videos.

Inspired by Explainable Artificial Intelligence (XAI) [207], XSec represents a novel paradigm in security research, aiming to enhance transparency, understanding, and trust in security mechanisms. It achieves this by providing clear explanations for their behavior, decisions, and vulnerabilities [208]. XSec’s application in video security systems marks a promising research area that integrates cutting-edge technologies to boost privacy and trust significantly. For instance, an XSec approach could be incorporated into security systems to explain the reasons behind the encryption of specific video feeds and control decisions and the effects of various security measures on the integrity of video data, thereby achieving the balance between security and transparency. However, XSec must be seamlessly integrated into current security workflows without causing disruptions or exposing sensitive data through explanations.

Video data often contains sensitive contextual information beyond faces, such as location, activities, or interactions. Preserving privacy while retaining this contextual information presents an ongoing research challenge. Simply blurring faces or other identifying features may prove insufficient to protect privacy, particularly with advancements in Facial Recognition Technology (FRT) [209]. Therefore, further research is necessary to explore more robust methods, such as Privacy-Enhancing Technologies (PETs), that maintain privacy while retaining the usefulness of the video data. PETs encompass various strategies, including homomorphic encryption, Secure Multiparty Computation (SMPC), Federated Learning, Trusted Execution Environments (TEEs), and Differential Privacy (DP), which minimise the collection, processing, and storage of personal data while reducing the risk of privacy breaches. Integrating these technologies will enable organisations to implement data protection measures by design and default, aligning with recommendations from General Data Protection Regulation (GDPR) while respecting privacy rights.

QSC, also known as Post-Quantum Cryptography (PQC) [210], involves designing and implementing cryptographic algorithms that are resistant to attacks from quantum computers. While some quantum-safe algorithms, such as lattice-based cryptography and hash-based signatures have been proposed [211]. Their practicality and efficiency in securing video data require a thorough investigation. Research is needed to explore the impact of quantum computing on video data and develop and optimise quantum-safe algorithms and protocols to protect videos against future quantum attacks.

The evolution of technologies like AI and IoT necessitates human-centric security approaches to safeguard video data. Users should ideally maintain good cyber hygiene [212] to protect data from hackers, yet many exhibit poor habits, such as sharing passwords and personal information on social networks [213]. Hackers exploit this vulnerability, making user information their easiest target. Thus, there is a need to recognise the importance of human-centric security in organisational culture, fostering collective responsibility. Traditional cybersecurity training focuses on dos and don’ts. In contrast, the human-centric approach delves into behaviour-based training, helping users understand the rationale behind security measures. Strategies like user authentication and security awareness programs are vital for building resilience against cyberattacks.

Increasing regulatory obligations and government controls over privacy and data localisation can significantly impact video data security. Adherence to regulations like GDPR (General Data Protection Regulation) [214], AI Act [215], CCPA (California Consumer Privacy Act) [216], HIPAA (Health Insurance Portability and Accountability Act) [217], and ISO (International Organisation for Standardisation)/IEC (International Electrotechnical Commission) 27001 [218] are crucial to avoid legal issues and maintain trust, as they impose requirements for transparency, consent, data protection, and accountability. Thus, there is a pressing need to investigate the evolving impact of regulations and industry standards on video data security and to develop compliance frameworks tailored to the unique challenges of securing video data. By addressing these research problems, organisations can effectively navigate the evolving regulatory landscape and implement robust security measures to safeguard videos while ensuring compliance with regulations and industry standards.