SPECPATCH

• DOI: 10.1145/3548606.3560660
• Publication Date: 2022-11-07T11:41:28Z
• Authors: Hanqing Guo; Yuanda Wang; Nikolay Ivanov; Li Xiao; Qiben Yan
• Keywords: Smartwatch

In this paper, we propose SpecPatch, a human-in-the loop adversarial audio attack on automated speech recognition (ASR) systems. Existing attacker assumes that the users cannot notice audios, and hence allows successful delivery of crafted examples or perturbations. However, in practical scenario, intelligent voice-controlled systems (e.g., smartwatches, smart speakers, smartphones) have constant vigilance for suspicious voice, especially when they are delivering their voice commands. Once user is alerted by audio, intend to correct falsely-recognized commands interrupting audios giving more powerful overshadow malicious voice. This makes existing attacks ineffective typical scenario user's interaction coincide. To truly enable imperceptible robust handle possible arrival interruption, design uses sub-second patch signal deliver an command utilize periodical noises break down communication between ASR We analyze CTC (Connectionist Temporal Classification) loss forwarding backwarding process exploit weakness achieve our goal. Compared with attacks, extend impact length (i.e., target command) 287%. Furthermore, show achieves 100% success rate both over-the-line over-the-air scenarios amid intervention.