US 11,856,024 B2
Prohibiting voice attacks
Bruno dos Santos Silva, McKinney, TX (US); Cheng-Ta Lee, Chamblee, GA (US); Howard Ruihua Zhao, Tucson, AZ (US); Socheat Sou, Ann Arbor, MI (US); James Edward Damgar, Pima, AZ (US); and George Patrick Carper, Tucson, AZ (US)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by International Business Machines Corporation, Armonk, NY (US)
Filed on Jun. 18, 2021, as Appl. No. 17/351,310.
Prior Publication US 2022/0407886 A1, Dec. 22, 2022
Int. Cl. H04L 9/40 (2022.01); G10L 21/02 (2013.01); G10L 15/10 (2006.01); G10L 15/22 (2006.01); G06F 16/22 (2019.01)
CPC H04L 63/1483 (2013.01) [G06F 16/2255 (2019.01); G10L 15/10 (2013.01); G10L 15/22 (2013.01); G10L 21/02 (2013.01); H04L 63/1416 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method comprising:
in response to receiving a voice input from a source, determining, using a predetermined filter including an allowlist, by one or more processors, that the voice input does not match any corresponding entry of the predetermined filter;
routing, by one or more processors, the voice input to an adversarial pipeline for processing by:
identifying an adversarial example of the voice input using a predetermined connectionist temporal classification (CTC) method;
generating a configurable distorted adversarial example using the adversarial example identified; and
in response to a user reply, injecting the configurable distorted adversarial example as noise into a voice stream of the user reply in real-time to alter the voice stream; and
routing, by one or more processors, the altered voice stream to the source.