Reproduce “Pmeta-TLA: Backdoor Attacks for Speech Classification Models via Meta-Learning with Timbre Leakage Attack” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yueming Huang, Wenhan Yao, Fen Xiao, Xiarun Chen, Weiping Wen

Researchers developed Pmeta-TLA, a backdoor attack strategy for speech classification models, by combining a stealthy frame-level timbre leakage attack (TLA) with a meta-learning framework and Projected Conflicting Gradients (PCGrad). This approach achieved high attack success rates (ASR) with reduced poisoning rates (PN) and demonstrated robustness against five distinct defense mechanisms, while enabling efficient multi-backdoor injection.

Key insights

  • Frame-level timbre leakage within deep self-supervised features can create highly stealthy and natural-sounding backdoor triggers that are difficult for humans and AI defenses to detect.
  • Meta-learning confers a "learn-to-inject-backdoor" capability, allowing attackers to rapidly adapt to and inject new backdoors with minimal additional poisoned data and fine-tuning.
  • Projected Conflicting Gradients (PCGrad) can effectively mitigate negative interference between the clean task and multiple backdoor tasks, enabling the injection of numerous backdoors while maintaining benign accuracy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.