[Paper Review] A Neural Network-Based On-device Learning Anomaly Detector for Edge Devices

Mineto Tsukada, Masaaki Kondo+1 more|arXiv (Cornell University)|Jul 23, 2019

Anomaly Detection Techniques and Applications50 references4 citations

TL;DR

This paper proposes ONLAD, a neural network-based on-device anomaly detector using Online Sequential Extreme Learning Machine (OS-ELM) for fast, low-power sequential learning on edge devices. It enables real-time concept drift adaptation with 1.95x–6.58x faster training and 5.0x–25.4x lower power consumption than software baselines, achieving standalone, communication-free operation on FPGA-based edge platforms.

ABSTRACT

Semi-supervised anomaly detection is an approach to identify anomalies by learning the distribution of normal data. Backpropagation neural networks (i.e., BP-NNs) based approaches have recently drawn attention because of their good generalization capability. In a typical situation, BP-NN-based models are iteratively optimized in server machines with input data gathered from edge devices. However, (1) the iterative optimization often requires significant efforts to follow changes in the distribution of normal data (i.e., concept drift), and (2) data transfers between edge and server impose additional latency and energy consumption. To address these issues, we propose ONLAD and its IP core, named ONLAD Core. ONLAD is highly optimized to perform fast sequential learning to follow concept drift in less than one millisecond. ONLAD Core realizes on-device learning for edge devices at low power consumption, which realizes standalone execution where data transfers between edge and server are not required. Experiments show that ONLAD has favorable anomaly detection capability in an environment that simulates concept drift. Evaluations of ONLAD Core confirm that the training latency is 1.95x~6.58x faster than the other software implementations. Also, the runtime power consumption of ONLAD Core implemented on PYNQ-Z1 board, a small FPGA/CPU SoC platform, is 5.0x~25.4x lower than them.

Motivation & Objective

• Address the limitations of server-based backpropagation neural networks (BP-NNs) in edge anomaly detection, including high latency and communication overhead.
• Overcome concept drift challenges in edge environments by enabling fast, sequential on-device learning.
• Design a lightweight, low-power hardware accelerator (ONLAD Core) for standalone execution on resource-constrained edge devices.
• Achieve high inference and training efficiency with minimal energy consumption and low latency for real-time anomaly detection.
• Enable deployment of semi-supervised anomaly detection models on edge devices without relying on remote servers or data transfers.

Proposed method

• Propose ONLAD, a semi-supervised anomaly detection model based on an autoencoder using Online Sequential Extreme Learning Machine (OS-ELM) for fast sequential learning.
• Optimize OS-ELM for single-batch (batch size = 1) training to drastically reduce computational cost and enable millisecond-scale updates.
• Introduce a lightweight forgetting mechanism based on FP-ELM to efficiently adapt to concept drift with minimal computational overhead.
• Implement ONLAD Core as a hardware IP core targeting FPGA platforms (e.g., PYNQ-Z1) for low-power, on-device training and inference.
• Use high-level synthesis (HLS) and PYNQ framework to co-design CPU-FPGA architectures that support real-time model updates without external server involvement.
• Integrate the ONLAD Core into a system that performs both training and inference on-device, eliminating data transfer to remote servers.

Experimental results

Research questions

• RQ1 Can ONLAD achieve comparable anomaly detection performance to traditional BP-NNs while enabling on-device sequential learning?
• RQ2 How effectively does ONLAD adapt to concept drift in real-time edge environments?
• RQ3 To what extent does ONLAD Core improve training speed and energy efficiency compared to software-based BP-NN and FP-ELM implementations?
• RQ4 Can ONLAD Core be efficiently deployed on low-power FPGA platforms like PYNQ-Z1 with practical model sizes?
• RQ5 Does the proposed forgetting mechanism in ONLAD outperform existing methods in terms of speed and adaptability to concept drift?

Key findings

• ONLAD achieves comparable generalization capability to BP-NN-based models in semi-supervised anomaly detection, especially under concept drift conditions.
• ONLAD Core reduces training latency by 1.95x to 6.58x compared to software implementations of BP-NNs and FP-ELM.
• The runtime power consumption of ONLAD Core on PYNQ-Z1 is 5.0x to 25.4x lower than software baselines during continuous training.
• The proposed forgetting mechanism in ONLAD is 3.21x faster on average than FP-ELM, enabling efficient concept drift adaptation.
• ONLAD Core is successfully implemented on the PYNQ-Z1 FPGA/CPU SoC platform with practical model sizes and low resource utilization.
• The system enables standalone on-device execution, eliminating the need for data transfers to remote servers and reducing communication latency and security risks.