Grant: Discovery Grants Program - Individual

In the expanding world of AI, speech processing emerges prominently. Voiceprints, which can be used by deep neural networks, have become a key identifier. These networks, skilled at analyzing patterns from signals, have widespread speech applications. However, their strengths also expose them to vulnerabilities. Among them, adversarial and poisoning attacks stand out as the main threats. Adversarial attacks add a small and often non perceptible noise to the signal in order to force the classifier to generate an incorrect output. The threats posed by such attacks are significant to the security of AI-based systems, particularly since a subtle adversarial change, imperceptible to humans, can result in unexpected system behavior. A real-world example could be a bank for which client accounts are protected by a speaker verification system. An attacker could add a small noise in order to fool the system and gain unauthorized access to personal information of several clients. Another threat to deep neural networks comes in the form of a poisoning attacks also called trojan attacks. This method involves introducing new samples into a database to create a "backdoor," triggering specific responses from neural networks trained on that data. The attacker achieves this by superimposing a "trigger" into the signal, eliciting a predetermined reaction from the neural network. Since deep neural networks often rely on public databases due to the important volume of data needed to train them, attackers can introduce tainted samples, potentially infecting numerous systems. From a practical point of view, such attacks are simpler than attacking a concealed model. The primary challenge lies in detecting such attacks since triggers can hide within millions of samples, masking themselves as typical sounds like music or conversations. The objective of this research work is to develop defense strategies against these types of attacks. The research plan includes experiments with quantum computers to see if this new technology could help defend classifiers against attacks. This research will play a crucial role in identifying vulnerabilities and implementing robust defense mechanisms for speech processing neural networks, thereby increasing security and confidence in these systems. Regarding the possibility of using quantum computers in this area, understanding their strengths, weaknesses and distinctive capabilities is pivotal for the development of machine learning algorithms utilizing this technology. The research program will also train a new generation of experts in these new technologies.