Reverse engineering deep learning algorithms
Cybersecurity concerns surrounding edge Field Programmable Gate Arrays (FPGA) accelerators hosting deep neural network (DNN) architectures have become increasingly prominent due to the vulnerability of such systems to side-channel reverse engineering attacks (SCAs). In this report, we present inn...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2024
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/175058 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Cybersecurity concerns surrounding edge Field Programmable Gate Arrays (FPGA)
accelerators hosting deep neural network (DNN) architectures have become increasingly
prominent due to the vulnerability of such systems to side-channel reverse engineering
attacks (SCAs). In this report, we present innovative defense mechanisms aimed at
thwarting SCAs on edge FPGA accelerators. Our study comprises two key experiments:
(1) Scheduling Measures against Hardware Trojan SCAs and (2) Dummy Layer
Obfuscator.
We first performed strategic modifications to the VTA Convolution 2D script to introduce
variability in computation order without compromising mathematical equivalence. While
significant disruptions to attack data volumes and layer identification are observed, the
study highlights a trade-off between security measures and computational efficiency.
Next, we introduced a Dummy Layer Obfuscator by strategically inserting dummy
convolutional layers into the DNN architecture to obscure its structure. This approach
successfully hinders attackers' ability to discern critical parameters, albeit with certain
limitations in layer placement and type.
Our findings underscore the importance of integrating robust security measures into the
design of FPGA-based DNN accelerators to safeguard against potential threats and uphold
model confidentiality. While our proposed defenses demonstrate effectiveness in
thwarting side-channel attacks, they also incur additional computational overhead. Future
research endeavors should focus on mitigating these overheads while preserving the
security benefits of the proposed solutions. |
---|