The article focuses on the rise of Machine Learning as a Service (MLaaS) and its potential vulnerability to backdoor attacks due to advancements in Artificial Intelligence (AI). The paper introduces a methodology called ‘DynamicTrigger’ for carrying out dynamic backdoor attacks. These attacks use clever modifications to deceive speech recognition systems, making corrupted samples indistinguishable from clean ones. The method involves using fluctuating signal sampling rates and masking speaker identities through dynamic sound triggers. The study shows that DynamicTrigger is potent and stealthy, achieving high success rates during covert attacks while maintaining accuracy with non-poisoned datasets. The paper also highlights the increasing use of AI in finance and the associated risks.
Publication date: 4 Jan 2024
Project Page: N/A
Paper: https://arxiv.org/pdf/2401.01537
