FGAS: Fixed Decoder Network-Based Audio Steganography with Adversarial Perturbation Generation
Hides secret messages in audio with 10 dB better quality than SOTA.
Researchers from Shanghai Jiao Tong University and collaborators have introduced FGAS (Fixed Decoder Network-Based Audio Steganography with Adversarial Perturbation Generation), a novel approach that hides secret messages in audio files by embedding them as adversarial perturbations. Unlike traditional encoder-decoder architectures that suffer from high computational costs and poor anti-steganalysis, FGAS uses a lightweight fixed decoder network shared between sender and receiver. The sender optimizes adversarial noise to carry the message while keeping the audio perceptually and statistically similar to the original, significantly improving stealth.
Experimental results show FGAS achieves an average PSNR gain of over 10 dB compared to state-of-the-art methods, indicating much higher stego audio quality. It also demonstrates strong robustness against common audio processing attacks like compression or filtering. Under high-capacity embedding, FGAS achieves a classification error rate about 2% higher than current SOTA, meaning it's harder for steganalysis tools to detect. This marks a significant step forward for covert communication in the age of AI-generated audio.
- FGAS hides secret messages as adversarial perturbations in audio, not in encoder-decoder networks.
- Achieves 10+ dB PSNR improvement over state-of-the-art steganography methods.
- Under high payloads, boosts anti-steganalysis by 2% classification error rate vs. SOTA.
Why It Matters
Enables more secure, high-quality covert audio communication resistant to detection and processing attacks.