Membership Inference Attacks against Large Audio Language Models

1. 1Identify potential vulnerabilities in Large Audio Language Models to membership inference attacks
2. 2Develop a multi-modal blind baseline using textual, spectral, and prosodic features to evaluate MIA performance
3. 3Assess train/test distribution shifts and their impact on MIA performance
4. 4Implement countermeasures to mitigate MIA risks in LALMs

AI engineers and researchers working with audio language models can benefit from understanding the vulnerabilities of these models to membership inference attacks, which can inform their design and testing strategies

💡 Large Audio Language Models are vulnerable to membership inference attacks due to train/test distribution shifts induced by non-semantic information in audio data

🔊 Researchers find near-perfect train/test separability in speech datasets, highlighting vulnerabilities of Large Audio Language Models to membership inference attacks