Natural-Sound Triggers Enable Near-Perfect Speech Model Backdoors

Mon, 20 Jul 2026 14:31:24 +0000

Engineer — Learn: Academic research showing ordinary ambient sounds can backdoor speech recognition models at only 5% poisoning rate with no clean-accuracy drop — informs threat modeling for teams training or fine-tuning ASR models, but no specific product or actionable patch is involved.
SOC/IR — Skip
Leader — Skip

Speech-Recognition on CuraSec