The rapid advancement of deepfake technology poses significant challenges to athletic celebrities, where altered or falsified media can impact athletes’ reputations, fan engagement, and the integrity of match broadcasting. This paper proposes a novel framework for deepfake image recognition for athletic celebrities using locality sensitive hashing (LSH). LSH, an efficient technique for high-dimensional nearest neighbor searches, is employed to detect and differentiate deepfake images from authentic media. By extracting high-dimensional features from images and videos using convolutional neural networks (CNNs), LSH is applied to hash similar content into clusters for quick and accurate deepfake detection. The proposed method is tested on real-world dataset, showing promising results in terms of accuracy and computational efficiency. This research highlights the importance of integrating advanced hashing techniques like LSH in safeguarding the authenticity of digital content and provides insights into future directions for deepfake detection mechanisms.

1. Introduction

The advent of digital technology has transformed the competitive industry in unprecedented ways, enhancing fan engagement and broadening the reach of athletic competitions [1, 2]. However, this transformation has also brought about new challenges, particularly in the realm of media authenticity. Among the most pressing issues is the rise of deepfake technology, which enables the creation of hyper-realistic yet entirely fabricated images and videos [3, 4]. As these technologies become increasingly sophisticated and accessible, they pose a significant threat to the integrity of media and the reputations of athletes.

Deepfakes are generated using advanced artificial intelligence techniques, particularly deep learning algorithms [5, 6]. These algorithms can manipulate video footage and images to create convincing representations of events that never occurred or alter the actions of real athletes in ways that could mislead audiences [7, 8]. The implications of such manipulations are profound; they can lead to misinformation, harm the credibility of match organizations, and affect the public’s trust in media. Given the unique dynamics of the competitive industry, where image and reputation are paramount, the need for robust deepfake detection mechanisms is critical. Athletes, teams, and organizations must be equipped with tools to identify and mitigate the risks posed by deepfakes. This is especially crucial in an era where social media serves as a primary source of information and news for millions of fans worldwide [9–11] The rapid spread of misleading content can lead to detrimental outcomes, ranging from reputational damage to financial losses for athletes and organizations alike. Traditional detection methods for deepfakes have focused on analyzing inconsistencies within the media itself, such as detecting unnatural facial movements or anomalies in lighting [12, 13]. However, as deepfake technology evolves, these methods are becoming less effective. To combat this challenge, this paper explores the application of locality sensitive hashing (LSH) as a powerful tool for deepfake image recognition, specifically within the competitive industry. LSH is a technique designed to efficiently handle high-dimensional data [14, 15]. LSH enhances rapid similarity searches by grouping similar data points into the same hash bucket. Within the realm of deepfake detection, LSH is utilized to hash features derived from images and videos. This method allows for the swift identification of potential deepfakes by comparing their features to those of verified authentic media, effectively determining similarity based on their hash bucket alignment. This approach not only enhances the speed of detection but also improves accuracy, as it allows for the consideration of complex, high-dimensional feature sets that characterize deepfake content. In general, the innovations of this article are three-fold.

1.
This study introduces the innovative application of LSH in deepfake image recognition specifically tailored for the competitive industry. By leveraging LSH’s ability to efficiently map high-dimensional feature spaces, the proposed framework enables rapid and accurate identification of deepfake images, significantly improving detection speed compared with traditional methods.
2.
The research employs convolutional neural networks (CNNs) to extract intricate and nuanced features from images and videos. This approach enhances the quality of the input data for the LSH algorithm, allowing for more precise comparisons and improved detection accuracy, thus addressing the limitations of existing detection techniques.
3.
The paper presents a comprehensive experimental framework that evaluates the performance of the proposed LSH-based detection method against existing state-of-the-art techniques. By utilizing real-world datasets and conducting rigorous benchmarking, the study provides empirical evidence of the effectiveness and efficiency of the proposed approach, thereby establishing a new standard for deepfake detection in this context.

The rest of article is organized as follows. The literature review in Section 2 summarizes existing deepfake detection methods, their limitations, and the relevance of LSH and CNNs. In Section 3, we detail the proposed detection framework, including feature extraction step, hash index creation step, and similarity evaluation step. In Section 4, the effectiveness of the proposed method is evaluated and compared with state-of-the-art techniques. At last, in Section 5, we summarize the key insights and suggest the directions for further research.

2. Related Work

In this section, we summarize the current research outcomes associated with this paper with the following categories: Deepfake Technology and Detection Methods and Application of LSH and CNNs.

2.1. Deepfake Technology and Detection Methods

The foundational paper [16] presents FaceForensics++, a large dataset of manipulated facial images and videos, which has become a benchmark for deepfake detection research. The dataset includes facial forgeries created using four popular manipulation techniques: FaceSwap, Face2Face, Neural-Textures, and DeepFakes. The paper [17] presents a new method for detecting deepfakes by examining convolutional traces; these are artifacts that deep generative models leave behind during the image generation process. The authors observe that deepfake images exhibit specific inconsistencies in their frequency domain representations, as generative models tend to manipulate images differently from natural images captured by cameras. In [18], the authors propose a method to detect deepfake videos by identifying face warping artifacts. Deepfake generation often involves resizing or warping a subject’s face to fit another body, leading to geometric inconsistencies between the face and the surrounding area. The authors develop a detection algorithm that analyzes these spatial inconsistencies, focusing on the discrepancies in the shape and appearance of facial features. In [19], MesoNet is introduced as a lightweight and compact neural network designed for real-time deepfake detection. The model is specifically built to capture mesoscopic-level features, which are intermediate features between pixel-level and global features in images.

The paper [20] takes a unique approach to deepfake detection by using biological signals such as subtle blood flow patterns in the human face, which are difficult for generative models to replicate accurately. The authors illustrate that deepfakes can be detected by analyzing the photoplethysmographic (PPG) signals derived from facial videos. They demonstrate that identifying disruptions or the absence of these natural biological rhythms is key to recognizing deepfake content. In [21], the authors explore the use of capsule networks for deepfake detection, proposing a method known as Capsule-Forensics. Capsule networks, unlike traditional CNNs, are designed to capture spatial relationships between different parts of an image. This makes them particularly effective at detecting subtle manipulations in deepfake videos, where the global structure of the face might remain intact, but local manipulations create inconsistencies. The survey paper [22] provides a comprehensive overview of deepfake detection techniques, covering a wide range of methods from image manipulation forensics to deep learning approaches. The authors categorize the detection methods into three main groups: traditional image forensics, machine learning-based detection, and physiological signal-based methods. The survey also highlights the strengths and weaknesses of each approach, offering insights into which methods are most effective under different conditions, such as varying image resolutions and compression levels. In [23], FakeCatcher is introduced as a tool to detect deepfake videos by examining biological signals, specifically the subtle color variations in human skin due to blood circulation. These natural fluctuations are typically present in authentic videos but are often missing or inaccurately represented in synthetic deepfake videos. The system extracts physiological cues from the video and uses these signals as a proxy for determining the authenticity of the video content.

However, the abovementioned literature seldom considers the high time cost for deepfake image recognition, which render current research outcomes are not very suitable for the big data scenarios where massive images need to be considered and handled.

2.2. Application of LSH and CNNs

The paper [24] presents an interpretable deepfake detection method by incorporating frequency spatial transformers into CNN filter kernels. The novel architecture enhances the ability to detect manipulated facial images by analyzing frequency and spatial domains concurrently, leading to improved accuracy in detecting deepfakes in compressed videos. In [25], the authors explore the use of a CNN-based discriminator in a GAN-based structure to enhance the detection and localization of small forgeries in satellite images. The approach achieves 86% accuracy in detecting subtle manipulations, making it applicable in media forensics, including deepfake detection in low-resolution videos. The authors of [26] address the challenges of detecting deepfakes by incorporating adversarial learning techniques and LSH for efficient hashing of media features. It focuses on applying CNNs for detecting synthetic faces in fake media, highlighting the use of LSH for fast retrieval of manipulated content. To improve processing speed, the authors in [27] introduce a sensitive content detection framework that employs perceptual hashing and CNNs to identify deepfake media. By integrating LSH with deep learning techniques, this approach enables adaptive detection. This combination enhances both the speed and robustness of real-time media monitoring.

The comprehensive survey in [28] highlights the use of CNN-based models for detecting copy-move forgeries and deepfakes. It emphasizes the potential of integrating LSH for rapid comparison of image regions, facilitating faster and more efficient detection in multimedia forensics. Similarly, literature [29] presents a CNN-based deepfake detection system aimed at identifying fake video clips. The system utilizes a combination of CNNs and LSH for hashing video features to accelerate the detection process, demonstrating superior performance in deepfake recognition tasks. Likewise, the authors in [30] provide a broad overview of current deepfake detection techniques, including LSH-based methods for quickly identifying manipulated media. The use of CNNs for feature extraction is discussed, and the paper highlights the importance of using LSH to speed up detection in large-scale datasets. The paper [31] systematically reviews various deepfake detection techniques and focuses on the role of CNNs in classification. It also examines the use of LSH for efficient deepfake detection by retrieving manipulated media from large datasets, making it ideal for social media monitoring. To improve the recognition performances, the work in [32] combines CNN and vision transformer architectures for video hashing and retrieval. LSH is employed for efficient media tracing, which is crucial for deepfake detection in large-scale video datasets. This hybrid approach enhances accuracy and reduces computational load.

However, the abovementioned literature did not the concrete characteristics of deepfake image recognition issue in competitive industry, which makes them hard to be directly employed in current athletics-related application scenarios. In the next section, we will introduce a new deepfake image recognition method tailed for competitive industry.

3. Our Proposal: DIR_LSH

In summary, our proposed deepfake detection method DIR_LSH could be divided into the following five steps (see Figure 1). Next, we will introduce the details of each step one by one.

Details are in the caption following the image — **Figure 1**
Open in figure viewer PowerPoint

Framework of our proposed deepfake recognition method DIR_LSH.

3.1. Step 1: Feature Extraction Using CNNs

The first stage in LSH-based deepfake detection involves extracting relevant features from video frames or images using CNNs. The features extracted from CNNs can represent essential information about the content [33, 34], such as facial expressions, lighting patterns, and textures, that can distinguish real from fake media.

Given a video frame I, a CNN model F (·) is applied to extract a feature vector. The input frame I is processed through several convolutional layers as follows:

()

where W_l represents the weights of the lth layer, b_l is the bias term, and σ (·) is the activation function (ReLU or Sigmoid). The output of each convolutional layer represents feature maps x_l, which encode spatial patterns within the frame.

After passing through multiple convolutional and pooling layers, the final feature vector f_I = F(I) is obtained as follows:

()

where d is the dimensionality of the feature space, typically large (e.g., d = 4096).

The CNN transforms an input image into a high-dimensional feature vector as follows:

()

where H is the height, W is the width, and C is the number of channels (color depth) of the image.

3.2. Step 2: Hashing Feature Vectors With LSH

Once the feature vectors are extracted, LSH is employed to map these high-dimensional vectors into lower-dimensional hash buckets while preserving the relative proximity of similar vectors.

1.
Random projection: LSH uses random hyperplanes to partition the feature space. A random vector r_i is sampled from a Gaussian distribution N (0, 1). LSH employs random projections, specifically through the use of random hyperplanes, as a fundamental technique for efficiently approximating nearest neighbor search in high-dimensional data spaces. The use of random vectors sampled from a Gaussian distribution N (0, 1) is crucial because Gaussian projections are known to preserve the dot product between vectors, a property pivotal in maintaining the geometric structure of the data. When a dataset is projected onto these random hyperplanes, the relative distances between points (both angular and Euclidean) are approximately maintained, which is vital for tasks such as nearest neighbor searches.
2.
Each feature vector f_I is projected onto random vectors as follows:
()
where sign (·) outputs one if the result is positive, and 0 otherwise. This creates a binary hash code for each feature vector.
3.
The process is repeated for k different random vectors, producing a k-dimensional binary hash code as follows:
()
where (f_I) ∈ {0, 1}^k.

For each feature vector f_I, LSH generates a binary hash code H(f_I) using k random hyperplanes as follows:

()

The locality-sensitive property ensures that for two similar feature vectors f_I and f_J, their hash codes H(f_I) and H(f_J) will likely be similar.

()

3.3. Step 3: Constructing the Hash Table

The binary hash codes generated from LSH are stored in a hash table. After generating the hash codes using LSH, constructing the hash table follows the following steps.

1.
Initialization: first, initialize a hash table. The structure of the table will depend on the number of distinct hash codes expected, which is influenced by the number of bits used in each hash code and the hash functions applied. Each unique hash code corresponds to a bucket within the table.
2.
Hash code generation: for each item in the dataset, compute its hash code. This involves projecting the item’s feature vector onto the predefined random hyperplanes and assigning a binary value based on the side of the hyperplane where the item lands. The sequence of binary values forms the hash code.
3.
Insertion into hash table: use the generated hash code as a key to place the item into the corresponding bucket in the hash table. If a bucket for a particular hash code does not yet exist, create it. If it does exist, add the item to the bucket.
4.
Handling collisions: since multiple items might generate the same hash code, these items will share the same bucket. This is intentional in LSH, as it groups similar items together, reducing the search space for similarity queries.

Each hash code serves as a key, and the corresponding feature vectors (or video frames) are stored as values.

1.
For each frame I, its hash code H(f_I) is used as a key in the hash table :
()
2.
During detection, when a new frame I^′ is input into the system, its hash code H(f_I) is computed, and the system looks for matching or similar hash codes in the table as follows:
()

The hash table stores feature vectors based on their binary hash codes as follows:

()

Given a new query frame I^′, the system retrieves all frames I such that .

3.4. Step 4: Similarity Search and Approximate Nearest Neighbor Retrieval

LSH allows for efficient approximate nearest neighbor search. When a new frame is tested, its feature vector is extracted, hashed, compared to the hash table.

1.
Given a query frame I^′, compute its feature vector and hash code as follows:
()
2.
Search for similar hash codes in the table . If a match I is found, compute the Euclidean distance between the feature vectors and f_I as follows:
()
3.
If the distance is below a predefined threshold τ, the frame I^′ is flagged as similar or potentially manipulated.

The approximate nearest neighbor search retrieves frame I such that:

()

where τ is the similarity threshold.

3.5. Step 5: Deepfake Detection Decision

After identifying frames with similar hash codes and evaluating their feature vector distances, the system makes a decision about whether the input frame I′ is a deepfake. The major basis for this process is the similarity or difference between different image frames. In concrete, if their similarity is smaller than a predefined threshold, then the new image frame could be regarded as a possible deepfake image; otherwise, the new image frame could be considered as authentic image.

1.
If the distance is smaller than τ, classify the frame as a deepfake as follows:
()
2.
Otherwise, classify the frame as authentic as follows:
()
The detection decision is formalized as a binary classification problem as follows:
()

The LSH-based deepfake detection system provides both computational efficiency and accuracy for identifying manipulated media in large-scale datasets. This method is particularly useful in detecting deepfakes in real-time, high-volume applications such as social media.

Pseudocode of the above DIR_LSH method is presented in Algorithm 1.

Algorithm 1: LSH-based Deepfake Detection Algorithm.

Input: I: Input frame
: Hash table
τ: Similarity threshold
Output: Label of the frame: “deepfake” or “authentic”
begin
Step 1: Feature Extraction
Extract feature vector f_I from input frame I using a CNN:
f_I⟵F(I)
Step 2: Hash Code Generation
Generate LSH-based hash code H(f_I) from the feature vector:
H(f_I)⟵[h₁(f_I), h₂(f_I), ⋯, h_k(f_I)]
Store hash code in the hash table :
Step 3: Query Hash Table for Similar Frames
For a new query frame I^′, extract feature vector:
Compute its LSH-based hash code:
Retrieve matching frames from the hash table:
Step 4: Compute Similarity Scores
For each retrieved frame I:
Step 5: Deepfake Detection
Compare similarity distance with threshold τ:
if then return “deepfake”
end else
return “authentic”
end
end

4. Experiments

4.1. Experimental Configuration

To validate the feasibility of the proposed DIR_LSH method, a group of experiment are tested in this section, which is conducted on the WS-DREAM dataset [35]. The dataset contains 1,974,675 values corresponding to 339 users and 5825 services. In the scenario of deepfake image detection focused in this article, we can regard the dataset as 339 images (parameter N), each of which is constituted by 5825 frames or dimensions (parameter d).

To observe the advantages of DIR_LSH method, we compare it with three SOTA methods, that is, UserCF [36], SS-ICF [37], and RP-UCF [38]. Evaluation performance metrics include MAE for measuring the accuracy and time cost for measuring the efficiency. Experiments are conducted on a Microsoft laptop with 12th Gen Intel (R) Core (TM) i5-1235U (2.50 GHz) and 16.0 GB RAM. Software settings include Win 11 OS and Python 3.6.

4.2. Evaluation Results

Next, we observe the performances of the proposed DIR_LSH method from various perspectives. In concrete, we have conducted the following four groups of experiments.

4.2.1. Profile 1: Accuracy Comparison

Here, we measure the accuracy of each method by calculating the average difference between the real value and predicted value of every frame in each image. In other words, we use the MAE metric to measure the deepfake detection accuracy. Regarding the parameters used in this profile, N is varied from 100 to 300 and d is varied from 1000 to 5000, respectively. Regarding our DIR_LSH method, there are two parameters, that is, size of hash function set k is equal to 8 and size of hash table set K is equal to 10.

Evaluation results are reported in Figure 2, where subgraph (a) is to measure the accuracy of four methods with respect to the variation of parameter N (here, another parameter d is fixed and equal to 1000) while subgraph (b) is to measure the accuracy of four methods with respect to the variation of parameter d (here, another parameter N is fixed and equal to 100). Overall, the fluctuation trend of the MAE values of four methods is not obvious with the growth of parameters N and d. In addition, the MAE values of our DIR_LSH method are much lower than the SS-ICF and RP-UCF methods, which indicate a higher accuracy of the proposed deepfake image detection solution in our article. Moreover, DIR_LSH and UserCF achieve the close MAE values with respect to parameters N and d, which show a good performance of DIR_LSH since UserCF is the classic baseline method.

4.2.2. Profile 2: Efficiency Comparison

Time efficiency is a key metric to measure the overall performance of big data-driven decision-making solutions [39–42]. The number of candidate images in deepfake image detection problem is often large and, therefore, the time cost for detection is a significant metric for performance evaluation. Inspired by this observation, we evaluate the time costs of different methods with respect to the parameters N and d. Similar to Profile 1, parameter N is varied from 100 to 300 and d is varied from 1000 to 5000, respectively. In our DIR_LSH method, the size of hash function set k is equal to 8 and the size of hash table set K is equal to 10.

Evaluation results are shown in Figure 3 where subgraph (a) is to measure the time costs of four methods with respect to the variation of parameter N (here, another parameter d is fixed and equal to 1000) while subgraph (b) is to measure the time costs of four methods with respect to the variation of parameter d (here, another parameter N is fixed and equal to 100). From the two figures, we can see an approximate increment trend of the time costs of four methods with the growth of N and d, which is because more images or more dimensions of each image often require additional processing time. Another observation from these two figures is that our proposed DIR_LSH method only needs less time costs compared to the rest three methods. This is because LSH used in DIR_LSH method has been proven a time-efficient neighbor searching technique and especially suitable in massive data processing scenarios.

4.2.3. Profile 3: Accuracy of DIR_LSH w.r.t. Parameters

In DIR_LSH method, there are still another two parameters k and K, which mean the size of hash function set and hash table set, respectively. In this profile, we evaluate the influence of parameters k and K toward the accuracy of DIR_LSH . Here, another two parameters N and d are equal to 100 and 1000, respectively. Evaluation results are shown in Figure 4. One observation from Figure 4 is that the MAE of DIR_LSH increases when K grows from 2 to 10. This is because more neighboring frames are judged to be similar to the target frame for deepfake image detection and as a result, the detection accuracy declines accordingly. The second observation from Figure 4 is that the MAE of DIR_LSH declines when k grows from 2 to 8. This is because when k grows, the returned neighboring frames are more similar to the target frame for deepfake image detection. Accordingly, the accuracy of DIR_LSH is enhanced.

4.2.4. Profile 4: Efficiency of DIR_LSH w.r.t. Parameters

In our DIR_LSH, k and K are two factors that are related to the algorithm efficiency. In this profile, we evaluate the influence of parameters k and K toward the time cost of DIR_LSH. Here, similar to Profile 3, another two parameters N and d are equal to 100 and 1000, respectively. Evaluation results are shown in Figure 5. The first observation from Figure 5 is that the time cost of DIR_LSH increases when K grows from 2 to 10. This is due to the fact that more neighboring frames are judged to be similar to the target frame for deepfake image detection and as a result, the detection time cost increases accordingly. Another observation from Figure 5 is that the time cost of DIR_LSH grows when k rises from 2 to 8. We can explain it as follows: when k grows, the evaluation condition for detecting similar frames is becoming narrow; in this situation, DIR_LSH probably cannot return any qualified similar frames, which may result in repeated executions of algorithm as well as higher time cost.

4.3. Further Discussion

Although our proposal achieves good performances in terms of deepfake recognition efficiency and accuracy, there are still several limitations. First of all, although our proposed deepfake recognition method based on LSH can achieve certain privacy protection capability, it is beneficial to integrate more effective privacy-preserving techniques into our work, such as unlearning [43], federated learning [44–46], and generalization technique [47]. In addition, when the image data for deepfake detection are massive, the so-called big data challenge is inevitable; in this situation, we need to recruit more time-efficient data processing techniques like the ones in [48–50]to achieve fast detection goal.

While our DIR_LSH method may not consistently surpass all other methods in deepfake image recognition accuracy, it generally performs better than most in terms of accuracy, as illustrated in Figure 2. Our DIR_LSH method exhibits significantly higher time efficiency compared with other methods, demonstrating swift deepfake image recognition, particularly in a big data context. Consequently, we contend that DIR_LSH is the optimal choice when considering the overall performance metrics, which include both time efficiency and recognition accuracy.

5. Conclusion

In this paper, we proposed a novel approach to deepfake detection in the competitive industry using LSH combined with CNNs. The primary goal was to enhance the efficiency and accuracy of detecting manipulated media, particularly in scenarios where large datasets, such as videos, are analyzed in real-time. In concrete, our method utilizes CNNs to extract high-dimensional feature vectors from athletic celebrities images and videos, which are then processed by LSH to generate hash codes that allow for efficient nearest-neighbor searches in high-dimensional spaces. By mapping similar frames into the same hash buckets, our approach significantly speeds up the retrieval of matching frames, making it ideal for handling the large volumes of media generated in the competitive industry. Experimental results demonstrate that the proposed system achieves promising performance in terms of both accuracy and computational efficiency when compared to state-of-the-art methods.

In the future work, we will further improve our proposal in this paper by introducing more context factors such as time and location. In addition, how to minimize the false-positive and false-negative probability incurred by the inherent limitation of LSH technique adopted in this paper is still another research topic that calls for intensive study in the future work.

Conflicts of Interest

The authors declare no conflicts of interest.

Funding

The authors have nothing to report.

Open Research

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

References

1 Peng Y., Liu J., Long M., and Peng F., Fldatn: Black-Box Attack For Face Liveness Detection Based on Adversarial Transformation Network, International Journal of Intelligent Systems. (2024) 2024, no. 1, https://doi.org/10.1155/2024/8436216.
10.1155/2024/8436216
Web of Science® Google Scholar
2 Yu R., Xiang J., Shu N. et al., Real-World Image Deraining Using Model-Free Unsupervised Learning, International Journal of Intelligent Systems. (2024) 2024, no. 1, https://doi.org/10.1155/2024/7454928.
10.1155/2024/7454928
Web of Science® Google Scholar
3 Zhao Q., Xu H., Li J., Rajput F. A., and Qiao L., The Application of Artificial Intelligence in Alzheimer’s Research, Tsinghua Science and Technology. (2024) 29, no. 1, 13–33, https://doi.org/10.26599/tst.2023.9010037.
10.26599/tst.2023.9010037
CAS Web of Science® Google Scholar
4 Wang J., Shi M., Zhang X. et al., Incremental Data Stream Classification With Adaptive Multi-Task Multi-View Learning, Big Data Mining and Analytics. (2024) 7, no. 1, 87–106, https://doi.org/10.26599/bdma.2023.9020006.
10.26599/bdma.2023.9020006
Web of Science® Google Scholar
5 Fei F., Li S., Dai H., Hu C., Dou W., and Ni Q., A K-Anonymity Based Schema for Location Privacy Preservation, IEEE Transactions on Sustainable Computing. (2019) 4, no. 2, 156–167, https://doi.org/10.1109/tsusc.2017.2733018.
10.1109/TSUSC.2017.2733018
Web of Science® Google Scholar
6 Dou W., Tang W., Wu X. et al., An Insurance Theory Based Optimal Cyber-Insurance Contract against Moral Hazard, Information Sciences. (2020) 527, 576–589, https://doi.org/10.1016/j.ins.2018.12.051, 2-s2.0-85059670190.
10.1016/j.ins.2018.12.051
Web of Science® Google Scholar
7 Zhou Q., Zhou Z., Bao Z., Niu W., and Liu Y., Iin-Ffd: Intra-Inter Network for Face Forgery Detection, Tsinghua Science and Technology. (2024) 29, no. 6, 1839–1850, https://doi.org/10.26599/tst.2024.9010022.
10.26599/tst.2024.9010022
Web of Science® Google Scholar
8 Reddy G. V., Deepika K., Malliga L. et al., Human Action Recognition Using Difference of Gaussian and Difference of Wavelet, Big Data Mining and Analytics. (2023) 6, no. 3, 336–346, https://doi.org/10.26599/bdma.2022.9020040.
10.26599/bdma.2022.9020040
Web of Science® Google Scholar
9 Wang F. et al., Ce-Rcfr: Robust Counterfactual Regression for Consensus-Enabled Treatment Effect Estimation, Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, ACM, 3013–3023.
Google Scholar
10 Wu T., Dou W., Wu F., Tang S., Hu C., and Chen J., A Deployment Optimization Scheme over Multimedia Big Data for Large-Scale Media Streaming Application, ACM Transactions on Multimedia Computing, Communications, and Applications. (2016) 12, no. 5s, 1–23, https://doi.org/10.1145/2983642, 2-s2.0-84994495804.
10.1145/2983642
Web of Science® Google Scholar
11 Liu W., Chen C., Liao X. et al., Learning Accurate and Bidirectional Transformation Via Dynamic Embedding Transportation for Cross-Domain Recommendation, Proceedings of the AAAI Conference on Artificial Intelligence. (2024) 38, no. 8, 8815–8823, https://doi.org/10.1609/aaai.v38i8.28728.
10.1609/aaai.v38i8.28728
Google Scholar
12 Yang B., Yang Y., Li Q. et al., Classification of Medical Image Notes for Image Labeling by Using Minbert, Tsinghua Science and Technology. (2023) 28, no. 4, 613–627, https://doi.org/10.26599/tst.2022.9010012.
10.26599/tst.2022.9010012
CAS Web of Science® Google Scholar
13 Yue P., Chen B., and Fu Z., Local Region Frequency Guided Dynamic Inconsistency Network for Deepfake Video Detection, Big Data Mining and Analytics. (2024) 7, no. 3, 889–904, https://doi.org/10.26599/bdma.2024.9020030.
10.26599/BDMA.2024.9020030
Web of Science® Google Scholar
14 Qi L., Wang R., Hu C., Li S., He Q., and Xu X., Time-Aware Dis-Tributed Service Recommendation With Privacy-Preservation, Information Sciences. (2019) 480, 354–364, https://doi.org/10.1016/j.ins.2018.11.030, 2-s2.0-85059185548.
10.1016/j.ins.2018.11.030
Google Scholar
15 Hu C., Fan W., Zen E. et al., Digital Twin-Assisted Real-Time Traffic Data Prediction Method for 5g-Enabled Internet of Vehicles, IEEE Transactions on Industrial Informatics. (2022) 18, no. 4, 2021–2032.
Web of Science® Google Scholar
16 Rossler A. et al., Faceforensics++: Learning to Detect Manipulated Fa- Cial Images, Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019.
Google Scholar
17 Guarnera L., Giudice O., and Battiato S., Deepfake Detection by Analyz-Ing Convolutional Traces, Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020.
Google Scholar
18 Li Y., Exposing Deepfake Videos by Detecting Face Warping Artifacts, 2018, https://arxiv.org/abs/1811.00656.
Google Scholar
19 Afchar D. et al., Mesonet: A Compact Facial Video Forgery Detection Network, 2018 IEEE International Workshop on Information Foren- Sics and Security (WIFS), 2018.
Google Scholar
20 Patil K., Kale S., Dhokey J., and Gulhane A., Deepfake Detection Using Biological Signals, 2023, https://arxiv.org/pdf/2301.05819.
Google Scholar
21 Nguyen H. H., Yamagishi J., and Echizen I., Capsule-forensics Net- Works for Deepfake Detection, Handbook of Digital Face Manipu-lation and Detection: From DeepFakes to Morphing Attacks, 2022, Springer International Publishing, Cham, 275–301.
Google Scholar
22 Swathi P. and Sk S., Deepfake Creation and Detection: A Survey, 2021 Third International Conference on Inventive Research in Comput- Ing Applications (ICIRCA), 2021.
Google Scholar
23 Ciftci U. A., Demir I., and Yin L., Fakecatcher: Detection of Synthetic Portrait Videos Using Biological Signals, IEEE Transactions on Pattern Analysis and Machine Intelligence. (2020) https://doi.org/10.1109/tpami.2020.3009287.
10.1109/tpami.2020.3009287
PubMed Google Scholar
24 Luan T., Liang G., and Pei P., Interpretable Deepfake Detection Based on Frequency Spatial Transformer, International Journal of Emerging Technologies and Advanced Applications. (2024) 1, no. 2, 19–25, https://doi.org/10.62677/ijetaa.2402108.
10.62677/ijetaa.2402108
Google Scholar
25 Fouad M. M., Mostafa E. M., and Elshafey M. A., Detection and Localization Enhancement for Satellite Images with Small Forgeries Using Modified Gan-based Cnn Structure, International Journal of Advances in Intelligent Informatics. (2020) 6, no. 3, 278–289, https://doi.org/10.26555/ijain.v6i3.548.
10.26555/ijain.v6i3.548
Google Scholar
26 Wang Y., Adversarial Deep Learning on Digital Media Security and Forensics, 2021, University of British Columbia, Ph.D. dissertation.
Google Scholar
27 Povedano Álvarez D., Sandoval Orozco A. L., García-Miguel J. P., and García Villalba L. J., Learning Strategies for Sensitive Content Detection, Electronics. (2023) 12, no. 11, https://doi.org/10.3390/electronics12112496.
10.3390/electronics12112496
Google Scholar
28 Zanardelli M., Guerrini F., Leonardi R., and Adami N., Image Forgery Detection: A Survey of Recent Deep-Learning Approaches, Multimedia Tools and Applications. (2023) 82, no. 12, 17 521–617 566, https://doi.org/10.1007/s11042-022-13797-w.
10.1007/s11042-022-13797-w
Web of Science® Google Scholar
29 Badale A. et al., Deep Fake Detection Using Neural Networks, 15th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS). (2018) 2.
Google Scholar
30 Verdoliva L., Media Forensics and Deepfakes: An Overview, IEEE Journal of Selected Topics in Signal Processing. (2020) 14, no. 5, 910–932, https://doi.org/10.1109/jstsp.2020.3002101.
10.1109/jstsp.2020.3002101
Web of Science® Google Scholar
31 Rana M. S., Nobi M. N., Murali B., and Sung A. H., Deepfake Detection: A Systematic Literature Review, IEEE Access. (2022) 10, 25 494–525 513, https://doi.org/10.1109/access.2022.3154404.
10.1109/access.2022.3154404
Web of Science® Google Scholar
32 Pei P., Zhao X., Li J., Cao Y., and Lai X., Vision Transformer-Based Video Hashing Retrieval for Tracing the Source of Fake Videos, Security and Communication Networks. (2023) 2023, 1–16, https://doi.org/10.1155/2023/5349392.
10.1155/2023/5349392
Google Scholar
33 Zhou X., Yang Q., Zheng X. et al., Personalized Federated Learning With Model-Contrastive Learning for Multi-Modal User Modeling in Human-Centric Metaverse, IEEE Journal on Selected Areas in Communications. (Apr 2024) 42, no. 4, 817–831, https://doi.org/10.1109/jsac.2023.3345431.
10.1109/jsac.2023.3345431
Web of Science® Google Scholar
34 Chen Y., Shu T., Zhou X. et al., Graph Attention Network With Spatial-Temporal Clustering for Traffic Flow Forecasting in Intelligent Transportation System, IEEE Transactions on Intelligent Transportation Systems. (2023) 24, no. 8, 8727–8737, https://doi.org/10.1109/tits.2022.3208952.
10.1109/TITS.2022.3208952
Web of Science® Google Scholar
35 Zheng Z. and Lyu M. R., Ws-Dream: A Distributed Reliability As-Sessment Mechanism for Web Services, 2008 IEEE International Conference on Dependable Systems and Networks with FTCS and DCC (DSN), 2008, IEEE, 392–397.
Google Scholar
36 Kluver D., Ekstrand M. D., and Konstan J. A., Rating-Based Collabora-Tive Filtering: Algorithms and Evaluation, Social Information Access: Systems and Technologies, 2018, Springer, 344–390.
Google Scholar
37 Li D. et al., Recommender System Frontier Topics, Recommender Systems: Frontiers and Practices, 2023, Springer Nature, Singapore, 205–229.
Google Scholar
38 Zheng Z., Li X., Tang M., Xie F., and Lyu M. R., Web Service Qos Prediction Via Collaborative Filtering: A Survey, IEEE Transactions on Services Computing. (2022) 15, no. 4, 2455–2472, https://doi.org/10.1109/tsc.2020.2995571.
10.1109/tsc.2020.2995571
Web of Science® Google Scholar
39 Zhou X., Wu J., Liang W. et al., Reconstructed Graph Neural Network With Knowledge Distillation for Lightweight Anomaly Detection, IEEE Transactions on Neural Networks and Learning Systems. (2024) 35, no. 9, 11 817–911 828, https://doi.org/10.1109/tnnls.2024.3389714.
10.1109/tnnls.2024.3389714
Web of Science® Google Scholar
40 Qi L., Xu X., Wu X., Ni Q., Yuan Y., and Zhang X., Digital-Twin-Enabled 6g Mobile Network Video Streaming Using Mobile Crowdsourc-Ing, IEEE Journal on Selected Areas in Communications. (2023) 41, no. 10, 3161–3174, https://doi.org/10.1109/jsac.2023.3310077.
10.1109/jsac.2023.3310077
Google Scholar
41 Kevin I. et al., Federated Transfer Learning Based Cross-Domain Pre-Diction for Smart Manufacturing, IEEE Transactions on Industrial Informatics. (2021) 18, no. 6, 4088–4096.
Google Scholar
42 Liu W., Xu X., Wu L. et al., Intrusion Detection for Maritime Transportation Systems With Batch Federated Aggregation, IEEE Transactions on Intelligent Transportation Systems. (2022) 24, no. 2, 2503–2514.
Google Scholar
43 Chen C., Zhang Y., Li Y. et al., Post-Training Attribute Unlearning in Recommender Systems, ACM Transactions on Information Systems. (2024) 43, no. 1, 1–28, https://doi.org/10.1145/3701987.
10.1145/3701987
CAS Google Scholar
44 Liang W., Chen X., Huang S., Xiong G., Yan K., and Zhou X., Federal Learning Edge Network Based Sentiment Analysis Combating Global Covid-19, Computer Communications. (2023) 204, 33–42, https://doi.org/10.1016/j.comcom.2023.03.009.
10.1016/j.comcom.2023.03.009
PubMed Web of Science® Google Scholar
45 Xu X., Li H., Li Z., and Zhou X., Safe: Synergic Data Filtering for Federated Learning in Cloud-Edge Computing, IEEE Transactions on Industrial Informatics. (Feb 2023) 19, no. 2, 1655–1665, https://doi.org/10.1109/tii.2022.3195896.
10.1109/tii.2022.3195896
Web of Science® Google Scholar
46 Zhou X., Liang W., Kawai A., Fueda K., She J., and Wang K., Adaptive Segmentation Enhanced Asynchronous Federated Learning for Sustainable Intelligent Transportation Systems, IEEE Transactions on Intelligent Transportation Systems. (Jul 2024) 25, no. 7, 6658–6666, https://doi.org/10.1109/tits.2024.3362058.
10.1109/tits.2024.3362058
Web of Science® Google Scholar
47 Qi L., Liu Y., Zhang Y., Xu X., Bilal M., and Song H., Privacy-Aware Point-of-Interest Category Recommendation in Internet of Things, IEEE Internet of Things Journal. (2022) 9, no. 21, 21 398–421 408, https://doi.org/10.1109/jiot.2022.3181136.
10.1109/jiot.2022.3181136
Web of Science® Google Scholar
48 Liang W., Xie W., Zhou X., I-Kai Wang K., Ma J., and Jin Q., Bi-Dueling Dqn Enhanced Two-Stage Scheduling for Augmented Surveillance in Smart Ems, IEEE Transactions on Industrial Informatics. (2023) 19, no. 7, 8218–8228, https://doi.org/10.1109/tii.2022.3216295.
10.1109/TII.2022.3216295
Web of Science® Google Scholar
49 Zheng X., Zhou X., Liang W., and I-Kai Wang K., Multi-Task Correlation Constrained Topological Learning Toward Smart Prognostic and Health Management in Iot, IEEE Internet of Things Journal. (2024) 11, no. 24, 39487–39496, https://doi.org/10.1109/jiot.2024.3446551.
10.1109/JIOT.2024.3446551
Google Scholar
50 Gu R., Wang S., Dai H. et al., Fluid-Shuttle: Efficient Cloud Data Transmission Based on Serverless Computing Compression, 2024, IEEE/ACM Transactions on Networking.
Google Scholar

Citing Literature

All articles

Locality Sensitive Hashing-Based Deepfake Image Recognition for Athletic Celebrities

Abstract

1. Introduction

2. Related Work