Real-Time Deepfake Detection in the Real-World (2406.09398v1)

Abstract: Recent improvements in generative AI made synthesizing fake images easy; as they can be used to cause harm, it is crucial to develop accurate techniques to identify them. This paper introduces "Locally Aware Deepfake Detection Algorithm" (LaDeDa), that accepts a single 9x9 image patch and outputs its deepfake score. The image deepfake score is the pooled score of its patches. With merely patch-level information, LaDeDa significantly improves over the state-of-the-art, achieving around 99% mAP on current benchmarks. Owing to the patch-level structure of LaDeDa, we hypothesize that the generation artifacts can be detected by a simple model. We therefore distill LaDeDa into Tiny-LaDeDa, a highly efficient model consisting of only 4 convolutional layers. Remarkably, Tiny-LaDeDa has 375x fewer FLOPs and is 10,000x more parameter-efficient than LaDeDa, allowing it to run efficiently on edge devices with a minor decrease in accuracy. These almost-perfect scores raise the question: is the task of deepfake detection close to being solved? Perhaps surprisingly, our investigation reveals that current training protocols prevent methods from generalizing to real-world deepfakes extracted from social media. To address this issue, we introduce WildRF, a new deepfake detection dataset curated from several popular social networks. Our method achieves the top performance of 93.7% mAP on WildRF, however the large gap from perfect accuracy shows that reliable real-world deepfake detection is still unsolved.

• The paper introduces LaDeDa, a patch-based detection algorithm that leverages local features to achieve 99% mAP on standard benchmarks.
• It distills LaDeDa into Tiny-LaDeDa, reducing computational complexity by 375x FLOPs and 10,000x parameters with minimal accuracy loss.
• The study also presents WildRF, a real-world social media dataset, highlighting the ongoing challenges in generalizing deepfake detection.

Real-Time Deepfake Detection in the Real-World

The significant advancements in generative AI facilitating the synthesis of hyper-realistic fake images pose new and critical challenges for the detection of such content. The paper "Real-Time Deepfake Detection in the Real-World" authored by Bar Cavia, Eliahu Horwitz, Tal Reiss, and Yedid Hoshen from the Hebrew University of Jerusalem, addresses this pertinent issue by introducing a novel approach for robust and efficient deepfake detection.

Summary

The authors present an innovative algorithm named "Locally Aware Deepfake Detection Algorithm" (LaDeDa), which demonstrates remarkable performance by leveraging local image features to identify deepfake content. LaDeDa scales down the deepfake detection task to a patch-level operation, accepting a solitary $9 \times 9$ image patch and outputting a patch-specific deepfake score. The compilation of these scores determines the overall image-level deepfake score. Notably, LaDeDa achieves a $99\%$ mean average precision (mAP) on existing benchmarks, significantly surpassing state-of-the-art (SoTA) methods.

Expanding on this, the authors distill LaDeDa into a more efficient model, Tiny-LaDeDa, maintaining high accuracy while drastically reducing computational complexity by $375 \times$ fewer FLOPs and $10,000 \times$ parameter reduction. The empirical results underscore the efficacy of Tiny-LaDeDa, marking a minor accuracy decrease for substantial gains in efficiency, making it suitable for deployment on edge devices.

Despite the promising results of LaDeDa and Tiny-LaDeDa on standard datasets, the authors bring to light a substantial gap in generalization when applying these models to real-world deepfakes sourced from social media platforms. To address this, they introduce WildRF, a comprehensive dataset curated from popular social networks such as Reddit, X (Twitter), and Facebook. Evaluation on WildRF shows that while LaDeDa achieves top performance with a $93.7\%$ mAP, the discrepancy from perfect accuracy suggests that the problem of real-world deepfake detection remains unsolved.

Key Contributions

1. Introduction of LaDeDa: A patch-based classifier that significantly enhances deepfake detection by focusing on local artifacts within small image patches, achieving state-of-the-art performance on standard benchmarks.
2. Distillation into Tiny-LaDeDa: Development of a highly efficient model that maintains competitive accuracy for real-time deepfake detection on edge devices, demonstrating a major reduction in computational resources.
3. WildRF Dataset: Creation of a realistic benchmark from social media data that encapsulates the diverse and complex nature of real-world deepfake scenarios, thereby providing a more accurate assessment of deepfake detection methods.

Detailed Analysis

The authors criticize current evaluation protocols for deepfake detection due to their inability to generalize well to real-world scenarios. They identify a critical flaw wherein standard datasets contain preprocessing discrepancies such as lossy JPEG compression for real images and lossless PNG compression for fake images. This discrepancy allows methods to artificially perform well on benchmarks that do not accurately simulate in-the-wild conditions.

LaDeDa circumvents this by using a ResNet50 variant that processes local patches of size $9 \times 9$, thus focusing on fine-grained image features rather than global semantics. The effectiveness of this method stems from vast supporting literature indicating that deepfake artifacts are typically low-level and localized. By averaging the patch-level scores, LaDeDa aggregates a comprehensive view that accurately classifies the image.

The innovation extends with Tiny-LaDeDa, using logit-based distillation to create a simplified version of the model. Tiny-LaDeDa demonstrates superior efficiency in computational tasks, suitable for deployment in resource-constrained environments such as mobile and embedded systems, without large sacrifices in accuracy.

Implications and Future Directions

The research presented offers considerable implications for both theoretical understanding and practical applications:

• Theoretical Impact: It underscores the importance of focusing on local image features in deepfake detection and opens avenues for further exploration in this architectural paradigm.
• Practical Relevance: The deployment viability of Tiny-LaDeDa on edge devices presents a practical solution for real-time deepfake detection, highlighting an essential use case for mobile and IoT security systems.

Future developments may involve extending the WildRF dataset to include a wider variety of social media platforms and generative models to cover a broader spectrum of real-world variations. Additionally, improving the robustness of detection algorithms against adversarial attacks and exploring the interpretability of model decisions could be fruitful areas of research.

In conclusion, the strides made by LaDeDa and Tiny-LaDeDa mark significant progress in the field of deepfake detection. However, the findings from the WildRF dataset emphasize the ongoing challenges and the necessity for continuous improvement to achieve reliable detection in real-world applications. This research paves the way for more advanced and practical deepfake detection technologies, integral to combating misinformation and safeguarding digital integrity.