Converting Autoencoder Based Energy-Efficient and Secure DNN Inference on Edge Devices

Deep neural networks (DNNs) have become essential for computer vision tasks like image classification, object detection, and depth estimation. With the rise of embedded devices, there is a growing demand for lightweight and energy-efficient models. While DNNs outperform traditional machine learning methods, their deep architectures pose challenges regarding energy consumption, latency, and computational efficiency. Despite various optimization techniques, there is still room for improvement in making these networks more efficient. We introduce a novel approach to DNN compression using autoencoders, leveraging the idea that not all images require the same level of complexity for classification. Our method trains an autoencoder to transform complex images into simpler representations, enabling a more efficient DNN. To optimize this process, we incorporate intraclass clustering on complex datasets, minimizing reconstruction loss and improving performance. This allows for the selective elimination of higher DNN layers, ensuring that the model meets its Service Level Objective (SLO) targets for various edge devices like Raspberry Pi and Jetson Nano. We also repurpose feature extraction layers from a baseline DNN as encoder layers while designing a decoder that reconstructs features in a way that simplifies classification. Instead of replicating input features, the autoencoder generates a more easily classifiable representation, enhancing accuracy while reducing computational overhead. By balancing efficiency and performance, our approach outperforms existing techniques, achieving the desired inference latency without compromising accuracy. As AI adoption continues to expand, more DNNs are being trained and deployed on third-party platforms, increasing security risks. Models operating in untrusted environments are particularly vulnerable to adversarial attacks, where small input v perturbations can lead to incorrect predictions. We have thoroughly examined both white-box attacks, which have full access to model parameters, and black-box attacks, which operate without knowledge of the model’s internal structure. To address these threats, we developed a detection framework that identifies compromised or Trojan inputs by analyzing activation patterns, entropy values, and reconstruction loss. By leveraging autoencoders to examine intermediate feature representations, our system effectively differentiates between benign and adversarial inputs, enhancing model security. Our goal is to build DNNs that are not only efficient but also robust against adversarial threats, ensuring both performance and security for real-world AI applications.

Author dc:creator

• Mahmud, Hasanul

Contributors dc:contributor

• Prasad, Sushil
• Lama, Palden
• Desai, Kevin
• Wang, Wei
• Xie, Mimi
• Prevost, Jeff

• DNN Optimization
• Edge Computing
• Energy Efficient
• Secure DNN
• Optimization
• Computational efficiency
• Service Level Objective

Language dc:language

eng

Identifier

9798314891919

OAI identifier oai:identifier

oai:tdl-ir.tdl.org:20.500.12588/7326