Harnessing ML Privacy by Design Through Crossbar Array Non-idealities

Conference · Mon Jun 10 04:00:00 EDT 2024

OSTI ID:2426423

Islam, Md Shohidul ^[1]; Dutta, Sankha B. ^[1]; Marquez, Andres ^[1]; Alouani, Ihsen ^[2]; KHASAWNEH, KHALED N. ^[3]

---

1. BATTELLE (PACIFIC NW LAB)
2. Queen's University Belfast
3. George Mason Universitiy

Deep Neural Networks (DNNs), handling computeand data-intensive tasks, often utilize accelerators like Resistiveswitching Random-access Memory (RRAM) crossbar for energyefficient in-memory computation. Despite RRAM’s inherent nonidealities causing deviations in DNN output, this study transforms the weakness into strength. By leveraging RRAM non-idealities, the research enhances privacy protection against Membership Inference Attacks (MIAs), which reveal private information from training data. RRAM non-idealities disrupt MIA features, increasing model robustness and revealing a privacy-accuracy tradeoff. Empirical results with four MIAs and DNNs trained on different datasets demonstrate significant privacy leakage reduction with a minor accuracy drop (e.g., up to 2.8% for ResNet-18 with CIFAR-100).

• Research Organization: Pacific Northwest National Laboratory (PNNL), Richland, WA (United States)
• Sponsoring Organization: USDOE
• DOE Contract Number: AC05-76RL01830
• OSTI ID: 2426423
• Report Number(s): PNNL-SA-194845
• Country of Publication: United States
• Language: English

Related Subjects

artificial intelligence