You Are the Trigger: Speaker-Activated Backdoors
Existing audio backdoors require an attacker to introduce triggers at inference time.
ProofPaper ↗
Key points
- Speech models are increasingly used as an interface to AI systems, from transcription tools to voice agents.
- We introduce the Speaker-Activated Backdoor (SAB) setting, in which a target speaker's natural voice serves as the trigger.
- We develop a two-phase fine-tuning method that first makes speaker identity more accessible and then binds the target speaker to an attacker-chosen payload.
- These results establish speaker identity as a practical backdoor trigger that requires no inference-time attacker access and is invisible at the transcript layer.
Sources (1)
- [1]You Are the Trigger: Speaker-Activated BackdoorsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 07:43 PM
Existing audio backdoors require an attacker to introduce triggers at inference time.
Speech models are increasingly used as an interface to AI systems, from transcription tools to voice agents.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 19, 2026How Deepgram optimized voice AI at scale with AWS Trainium
- Oct 19, 2026Push-based KV transfer to improve TTFT and TPOT for disaggregated inference in vLLM
- Oct 19, 2026NIOCIS: Typed-Decision Next-Hop Routing with Node-Conditioned Option Spaces for Tool-Use Agents
- Oct 19, 2026Standing on the Shoulders of Giants: Enhancing LLM Agents for Automated Fact-Checking with Human Expert Skills
- Oct 19, 2026From Black Box to Controller: Inference-Time Steering of LLM via Feature Amplification
- Oct 18, 2026Google Open-Sources Harness2, a Recursive Agent Harnessing Framework