Quantized models remain vulnerable to privacy attacks

Practical quantization levels have limited impact on membership privacy risks.

Quantization is a technique used to reduce model size for efficient deployment of large-scale deep neural networks (DNNs).

Practical quantization levels of 44-88 bits leave quantized models largely as vulnerable to membership inference attacks (MIAs) as full-precision counterparts.

Aggressive low-bit quantization like 22-bit can provide meaningful privacy gains, but this comes at the cost of substantial utility degradation.

Membership inference attacks (MIAs) are used to evaluate training data privacy risks.

Quantization is a technique for reducing model size.

The study evaluated eight state-of-the-art MIA attacks across six representative quantization methods.

Key facts

  • Quantized models at 44-88 bits remain largely as vulnerable to membership inference attacks (MIAs) as full-precision models.
  • Low-bit quantization of 22-bit can provide privacy gains but causes substantial utility degradation.
  • Existing privacy defenses for full-precision DNNs provide limited protection for quantized models and increase utility degradation.

All sections

Edition