Source-linked AI summary
Continual Learning and Private Unlearning
Bo Liu, Qiang Liu, Peter Stone
TL;DR
The paper addresses how an agent can learn tasks continually yet later forget selected tasks privately without degrading retained knowledge. It formalizes CLPU, proposes CLPU-DER++ as an exact initial solution, and introduces benchmarks and metrics for evaluation. The method achieves exact unlearning and good continual-learning performance, while using additional model space.
Problem
Agents may need to learn tasks temporarily and later forget them immediately and privately, without revealing information about the task or degrading other learned knowledge.
Method
The paper formalizes CLPU and proposes CLPU-DER++, combining dynamic architecture and rehearsal, with benchmark tasks and evaluation metrics.
Results
CLPU-DER++ achieves exact unlearning and good continual-learning performance, while trading off additional model space.
Takeaways & Limitations
CLPU-DER++ provides an initial solution for jointly evaluating continual knowledge transfer and private unlearning.
Takeaways & Limitations
The formulation assumes no dataset is supplied when a previously temporary task is later remembered or forgotten, and shared task loss functions are assumed in practice.
Abstract
from arXiv · showhide
As intelligent agents become autonomous over longer periods of time, they may eventually become lifelong counterparts to specific people. If so, it may be common for a user to want the agent to master a task temporarily but later on to forget the task due to privacy concerns. However enabling an agent to \emph{forget privately} what the user specified without degrading the rest of the learned knowledge is a challenging problem. With the aim of addressing this challenge, this paper formalizes this continual learning and private unlearning (CLPU) problem. The paper further introduces a straightforward but exactly private solution, CLPU-DER++, as the first step towards solving the CLPU problem, along with a set of carefully designed benchmark problems to evaluate the effectiveness of the proposed solution. The code is available at https://github.com/Cranial-XIX/Continual-Learning-Private-Unlearning.
1 INTRODUCTION
The paper formulates continual learning and private unlearning (CLPU) for agents that must learn tasks temporarily or permanently and later forget selected tasks without privacy leakage. It proposes CLPU-DER++ as an exact initial solution, evaluated with new benchmarks.
- Motivation: CL usually aims to preserve prior knowledge while adapting to new tasks, but forgetting can also be deliberately useful for abstraction and privacy.The paper motivates controlled forgetting when learned behaviors or knowledge create fairness, privacy, or security concerns.
- Motivation: Private unlearning must occur immediately and prevent both recovering information about the task and detecting that the task was previously learned.Natural catastrophic forgetting may be insufficient because it can be slow and may not remove evidence of prior learning.
- Problem formulation: CLPU lets users instruct an agent to remember tasks permanently, learn them temporarily, or forget them completely and privately.The paper distinguishes this setting from prior selective forgetting that degrades task performance but is not privacy-preserving.
- Approach: CLPU-DER++ combines dynamic architecture and rehearsal to provide a straightforward exact-unlearning solution for continual learning.The method is presented as an initial solution rather than a complete resolution of the trade-offs among transfer, model space, and privacy.
- Evaluation: The proposed benchmarks evaluate both continual-learning effectiveness and private-unlearning behavior.The contribution is framed as a formalized problem, an initial exact-unlearning method, and benchmark tasks for evaluation.
2 RELATED WORK
The related-work discussion places CLPU between continual learning, which preserves knowledge across sequential tasks, and machine unlearning, which removes requested training influence.
- Connection to CLPU: CLPU connects these two areas by combining sequential learning with user-requested private forgetting.The paper summarizes the relationship between CLPU and prior literature in Figure 1.
- Continual learning: Continual learning studies sequential task acquisition when previous data are generally unavailable, with methods designed to mitigate forgetting.The literature spans computer vision, reinforcement learning, natural language processing, and robotics.
- Machine unlearning: Machine unlearning removes the influence of a requested subset of training data, often avoiding full retraining through data partitioning or stored model structure.Its goal differs from continual learning because it focuses primarily on removing specified training influence from a learned model.
3 BACKGROUND
The background defines continual learning as sequential supervised-task learning without access to earlier data, and machine unlearning as removing a forget set while preserving retained-task performance.
- Continual learning: In continual learning, an agent sequentially learns K supervised tasks and ultimately seeks low loss across their test datasets despite losing access to earlier training data.The model is parameterized by θ, and task performance is evaluated through task-specific losses.
- Continual learning: The setup assumes shared task loss functions in practice and replaces testing loss with training loss because test data are unavailable during training.This assumption is explicitly stated for the remainder of the paper.
- Continual learning: Regularization-based continual learning preserves prior knowledge by constraining updates near earlier parameters, while rehearsal stores a small episodic memory of exemplars.Rehearsal methods trade off learning new knowledge against preserving old knowledge, including through saved logits in DER++.
- Machine unlearning: Machine unlearning receives a forget set Df and retained set Dr, returning a model that contains no information about Df while performing well on Dr.Exact unlearning requires the resulting model distribution to match training directly on Dr.
- Machine unlearning: The unlearning formulation is restrictive, and δ-unlearning relaxes exact matching while permitting asymmetric output distributions.The cited background notes that exact unlearning can be difficult to achieve in practice.
- Machine unlearning: Unlearning can be evaluated through an output function when an adversary observes predictions rather than model parameters, and exact removal alone is insufficient without retained-data performance.The paper also distinguishes exact unlearning from approximate unlearning and from differential privacy.
4 PROBLEM AND METHOD
CLPU formalizes sequential task learning with user instructions to retain, temporarily learn, or privately forget tasks. CLPU-DER++ uses isolated temporary models and episodic memories so forgotten tasks can be removed exactly while other knowledge remains in the main model.
- CLPU formulation: CLPU agents process sequential requests specifying whether each task should be permanently learned, temporarily learned, or exactly forgotten.The task-status dictionary records retained or temporary tasks and removes a task upon an unlearning request.
- CLPU formulation: During ordinary learning, the agent must preserve performance on previously learned tasks, while unlearning requires model parameters indistinguishable from those produced without the forgotten task.The formulation uses a distribution distance D for the unlearning condition and continues learning over the remaining tasks.
- CLPU formulation: CLPU differs from conventional continual learning and machine unlearning by combining knowledge transfer with task-specific private removal under online task ordering and limited access to prior data.The setting requires maintaining knowledge unrelated to the forgotten task rather than merely degrading performance on that task.
- CLPU-DER++: CLPU-DER++ adapts DER++ using a main model, isolated temporary networks for temporary tasks, and episodic memories for subsequent updates.Temporary learning initializes an isolated model from the main model, trains it on the task data, and stores its episodic memory.
- CLPU-DER++: Forgetting a temporary task removes its temporary network and corresponding episodic memory without changing the main model, achieving exact unlearning by construction.The method stores a full extra model for each temporary task, trading memory for privacy.
5 EXPERIMENTAL RESULTS
The experiments adapt four conventional continual-learning benchmarks to CLPU, evaluate continual-learning and private-unlearning behavior with complementary metrics, and compare CLPU-DER++ against adapted baselines. CLPU-DER++ provides the strongest private-unlearning results while remaining close to the best continual-learning performance.
- 5.1 CLPU Experiment Setup: The study compares CLPU-DER++ with sequential learning, independent learning, Elastic Weight Consolidation, Experience Replay, DER++, and Learning with Selective Forgetting.The baselines are adapted to the CLPU setting, while CLPU-DER++ combines dynamic architecture and rehearsal approaches.
- 5.1 CLPU Experiment Setup: The evaluation adapts Perm-MNIST, Rot-MNIST, Split-CIFAR10, and Split-CIFAR100 into CLPU benchmarks with sequential user requests for learning, temporary learning, retention, and forgetting.The benchmarks use five sequential tasks for Rot-MNIST and Perm-MNIST, with request sequences that include both ordinary learning and unlearning.
- 5.2 Evaluation Metrics: ACC measures final average accuracy and FM measures average performance drop on tasks that remain in the final task-status dictionary.Both metrics evaluate continual-learning performance after processing the request sequence, using holdout testing data.
- 5.2 Evaluation Metrics: Private unlearning is evaluated by comparing post-unlearning output distributions with distributions from models that would not have learned the forgotten task, using IJSD, AJSD, JS-ratio, and IRR.IJSD compares models within the unlearning group, while AJSD compares unlearning-group models with models trained without the task; JS-ratio and IRR summarize these distances.
- 5.3 Results: CLPU-DER++ achieves the best JS-ratio and IRR across the benchmark comparisons, whereas baselines show high JS-ratio and very low IRR.The reported baseline pattern indicates that unlearning with those methods reveals that the model previously learned the unlearned task.
- 5.3 Results: CLPU-DER++ finishes a close second on continual-learning performance because merging a temporary network uses knowledge distillation, which may not fully recover the original model’s performance.Initializing the temporary network from the main model performs better than initializing it from scratch.
6 CONCLUSION AND FUTURE WORK
The paper formalizes continual learning and private unlearning, introduces CLPU-DER++ as an exact unlearning method, and provides metrics and adapted benchmarks for evaluation. It identifies extending to δ-unlearning, reducing space complexity, understanding optimal trade-offs, and studying task relationships as future directions.
- 6 CONCLUSION AND FUTURE WORK: The paper formalizes the continual learning and private unlearning problem and introduces CLPU-DER++ as a straightforward but exact unlearning method.It also introduces novel metrics and adapted benchmark problems for evaluating CLPU methods.
- 6 CONCLUSION AND FUTURE WORK: CLPU-DER++ is presented as an initial solution with exact privacy and good knowledge-transfer ability.The paper proposes extending it to δ-unlearning while saving fewer models to reduce space complexity.
- 6 CONCLUSION AND FUTURE WORK: Future work includes characterizing optimal CLPU performance and studying how task relationships affect continual learning and private unlearning.The paper notes that the three objectives may not be simultaneously optimizable and that similar tasks may improve continual learning while making private unlearning more difficult.