[공지사항]을 빙자한 안부와 근황 
Show more

Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

SHIELD: A Secure and Highly Enhanced Integrated Learning for Robust Deepfake Detection against Adversarial Attacks

Created by
  • Haebom

Author

Kutub Uddin, Awais Khan, Muhammad Umar Farooq, Khalid Malik

Outline

In this paper, we propose SHIELD, a novel collaborative learning method to address the vulnerability of deepfake audio detection. We experimentally show that existing deepfake audio detection methods are vulnerable to anti-forensics (AF) attacks based on generative adversarial networks, and design a collaborative learning framework that integrates a defensive generative model (DF) to defend against them. SHIELD uses a triplet model to capture the correlation between real and AF attack audio, and real and attack audio generated using an auxiliary generative model. It demonstrates strong performance on various generative models on ASVspoof2019, In-the-Wild, and HalfTruth datasets, and effectively mitigates the degradation of detection accuracy caused by AF attacks.

Takeaways, Limitations

•
Takeaways:
◦
Introducing SHIELD, a new defense technique that effectively counters anti-forensic attacks on deepfake audio.
◦
Improving robustness against AF attacks through collaborative learning using auxiliary generative models.
◦
Performance validation through experiments on various datasets and generative models.
•
Limitations:
◦
Further research is needed on the generalization performance of the proposed SHIELD. Further evaluation of its resistance to various types of AF attacks may be needed.
◦
Performance evaluation in real-world environments may be limited. Testing on a variety of real-world deepfake audio data may be required.
◦
Consideration should be given to computational cost and complexity. Additional research may be needed to improve the effectiveness of SHIELD.
👍
Made with Slashpage