Table 2 An overview of the query-based attacks against ASR
From: Towards the universal defense for query-based audio adversarial attacks on speech recognition system
Attack | Task | Attack method | Attack model | Target | M or D | Avg.Queries | SRoA(%) |
|---|---|---|---|---|---|---|---|
CS Yuan et al. (2018) | ASR | GD | Kaldi-Aspire | Play music. Open the front door. Turn off the light. | M | \(\sim\)300 | 100 |
DS Carlini and Wagner (2018) | ASR | GD | DeepSpeech | Okay google browse to evil dot com. | M & D | \(\sim\)1000 | 100 |
DW Chen et al. (2020) | ASR | Alt-M | APIs | Turn off The Light Take a picture. Call 911. | M | \(\sim\)150 | 100 |
DSG Taori et al. (2019) | ASR | GA & GE | DeepSpeech | Morning body. Ball charge. More they. | D | \(\sim\)150000 | 35 |
Foolgle Han et al. (2019) | ASR | GA | Google-API | – | D | – | 86 |
SGEA Wang et al. (2021) | ASR | SGE | DeepSpeech | Thank you. Hello world. Open the door. | D | \(\sim\)78000 | 98 |
IRTA Qin et al. (2019) | ASR | Psy-M | Lingvo | Old will is a fine fellow but poor and helpless sin -ce missus rogers had her accident. | D | \(\sim\)5000 | 100 |
PHA Schönherr et al. (2018) | ASR | Psy-M | Kaldi-WSJ | Do not blame you. The command is planted. The cake is a lie. | M & D | \(\sim\)500 | 98 |
EPA Abdullah et al. (2021) | ASR | Psy-M | DeepSpeech and Wav2Letter | That is comparatively nothing. Talking later is beneath us. But there seemed no. | D | \(\sim\)1000 | 76 |
Occam Zheng et al. (2021) | ASR | Co-E | DeepSpeech and APIs | Call my wife. Navigate to my home. Open the door. | M & D | \(\sim\)30000 | 100 |
SirenAttack Du et al. (2020) | ASR | PSO | DeepSpeech | Read last sms from boss. Call the police for help. | D | \(\sim\)1000 | 100 |
MOGA-Attack Khare et al. (2018) | ASR | Mul-Obj GO | DeepSpeech and Kaldi | A cat. All of these. That i love you. | D | - | * |