FiCABU: A Fisher-Based, Context-Adaptive Machine Unlearning Processor for Edge AI

TL;DR

FiCABU addresses the need for privacy-driven unlearning directly on edge devices by combining Context-Adaptive Unlearning with Balanced Dampening in a SW–HW co-design. It starts edits from back-end layers, uses depth-aware dampening, and stops once the forget objective is reached, all while caching activations to enable efficient partial inferences. Implemented as a three-stage GEMM-centric streaming processor with dedicated FIMD and Dampening IPs, FiCABU demonstrates near-SSD-quality retention and substantial computation/energy savings on CIFAR-20 and PinsFaceRecognition, including an INT8 hardware prototype. The results validate that on-device, back-end-first, depth-aware unlearning is practical for resource-constrained edge AI workloads.

Abstract

Machine unlearning, driven by privacy regulations and the "right to be forgotten", is increasingly needed at the edge, yet server-centric or retraining-heavy methods are impractical under tight computation and energy budgets. We present FiCABU (Fisher-based Context-Adaptive Balanced Unlearning), a software-hardware co-design that brings unlearning to edge AI processors. FiCABU combines (i) Context-Adaptive Unlearning, which begins edits from back-end layers and halts once the target forgetting is reached, with (ii) Balanced Dampening, which scales dampening strength by depth to preserve retain accuracy. These methods are realized in a full RTL design of a RISC-V edge AI processor that integrates two lightweight IPs for Fisher estimation and dampening into a GEMM-centric streaming pipeline, validated on an FPGA prototype and synthesized in 45 nm for power analysis. Across CIFAR-20 and PinsFaceRecognition with ResNet-18 and ViT, FiCABU achieves random-guess forget accuracy while matching the retraining-free Selective Synaptic Dampening (SSD) baseline on retain accuracy, reducing computation by up to 87.52 percent (ResNet-18) and 71.03 percent (ViT). On the INT8 hardware prototype, FiCABU further improves retain preservation and reduces energy to 6.48 percent (CIFAR-20) and 0.13 percent (PinsFaceRecognition) of the SSD baseline. In sum, FiCABU demonstrates that back-end-first, depth-aware unlearning can be made both practical and efficient for resource-constrained edge AI devices.

Figures (6)

• Figure 1: Comparison of server-based unlearning, which incurs communication and privacy costs, and adaptive on-device unlearning, which performs selective updates locally.
• Figure 2: Overview of FiCABU: context-adaptive unlearning, balanced dampening, and edge-oriented implementation.
• Figure 3: Layer-wise distribution of selected parameters for ResNet-18 (RN) and ViT, highlighting concentration in back-end layers.
• Figure 4: Baseline uniform scaling vs. proposed sigmoid-based profile; $S(l)$ is smaller at the back-end and larger at the front-end.
• Figure 5: Key hardware components of FiCABU: (a) FIMD module for diagonal Fisher estimation used by Context-Adaptive Unlearning, (b) Dampening module implementing Balanced Dampening updates, and (c) patch-level GEMM–FIMD–Dampening pipeline.