REVOLUTIONIZING TELECOM SECURITY AND RELIABILITY THROUGH DE-IDENTIFIED AND ROBUST COMMUNICATION

forthcoming

Privacy infringment is one of the most prevalent forms of cybercrime, accounting for about 72% of all cybercrimes alone. Such is also the case with speaker biometrics, which, if compromised over a call, can be manipulated to generate fake speech to impersonate a person. Such identity theft over voice calls can cause financial losses, cyber-bullying and fake hate speech generation. On the other hand, in parts of the world where the communication infrastructure is not robust, a person is more agitated over not being able to get his message across over a call due to channel breakage. In such cases, privacy concerns take a back seat; but that does not imply the absence of the threat. To that end, I propose a joint solution to these issues, that can be integrated with the existing communication infrastructure. The solution consists of two segments: far-end and near-end. The far-end segment shall handle the speaker de-identification task. Existing methods of speaker anonymization are limited in terms of achievable anonymity. Improvements can be made through cycle-consistent GAN to generate better synthetic speaker representation vectors. I plan on surpassing the existing benchmark in terms of lower automatic speaker verification (ASV) scores and good automatic speech recogntion (ASR) scores. Additionally, the entire segment would be brought online for integration with real-time communication systems. The near-end segment shall be handling the speech inpainting task. While existing techniques are able to fill the blanks in acoustic signals pertaining to music and background sounds quite well, they are ill-equipped for speech. The existing speech inpainting techniques chiefly rely on spectrogram-based approaches which result in unnatural reconstructions. I plan to study novel approaches of speech inpainting, including attention based formant imputation techniques and surpass the current benchmarks for successfully filling speech gaps in the range of 1600ms.