Replay Attacks: The Blind Spot in Audio Deepfake Detection

Groundbreaking research from Resemble AI and collaborators, entitled “Replay Attacks Against Audio Deepfake Detection,” has been accepted for presentation at the Interspeech 2025 conference, highlighting a new challenge in deepfake detection. The study reveals that replay attacks, where generative AI audio is played through speakers and re-recorded, can alter acoustic properties and make deepfakes appear authentic, posing a significant vulnerability to current detection methods. The research introduces ReplayDF, a comprehensive dataset designed to study and combat these attacks, featuring diverse conditions and real-world scenarios to enhance detection models. The study found that replay attacks significantly degrade the performance of existing detection models, but adaptive retraining using Room Impulse Responses (RIRs) can improve their robustness. Resemble AI's Detect platform is at the forefront of addressing these challenges, offering a state-of-the-art neural model for real-time deepfake audio detection, alongside comprehensive solutions like AI watermarking and multimodal protection. The research underscores Resemble AI's commitment to leading the evolution of generative AI safety by proactively mapping challenges and building defenses, ensuring digital authenticity in an increasingly synthetic media landscape.