Source-linked AI summary
Can Bad Teaching Induce Forgetting? Unlearning in Deep Networks using an Incompetent Teacher
Vikram S Chundawat, Ayush K Tarun, Murari Mandal, Mohan Kankanhalli
TL;DR
Machine unlearning must remove requested data from trained models while meeting privacy requirements and avoiding costly retraining. The paper uses competent and incompetent teachers to selectively transfer knowledge to a student, and reports effective forgetting across settings while introducing ZRF for retrained-model-free evaluation.
Problem
Data-deletion requirements create a need to remove information from trained ML models, while existing evaluation can depend on an expensive retrained reference model.
Method
A student initialized from the original model selectively learns accurate knowledge from a competent teacher on retain data and bad knowledge from an incompetent teacher on forget data.
Results
The method generalizes across class, subclass, and random-sample forgetting, with ZRF reaching 0.99 from 0.87 and membership-inference probability dropping to 0.002 from 0.982 for rocket-class samples.
Takeaways & Limitations
ZRF enables evaluation when a retrained model is unavailable or expensive, while the teacher-student method supports multiple deep-network modalities and application domains.
Takeaways & Limitations
ZRF uses test-set performance as a practical proxy for the ideal retrained-model value when the retrained model is unavailable.
Abstract
from arXiv · showhide
Machine unlearning has become an important area of research due to an increasing need for machine learning (ML) applications to comply with the emerging data privacy regulations. It facilitates the provision for removal of certain set or class of data from an already trained ML model without requiring retraining from scratch. Recently, several efforts have been put in to make unlearning to be effective and efficient. We propose a novel machine unlearning method by exploring the utility of competent and incompetent teachers in a student-teacher framework to induce forgetfulness. The knowledge from the competent and incompetent teachers is selectively transferred to the student to obtain a model that doesn't contain any information about the forget data. We experimentally show that this method generalizes well, is fast and effective. Furthermore, we introduce the zero retrain forgetting (ZRF) metric to evaluate any unlearning method. Unlike the existing unlearning metrics, the ZRF score does not depend on the availability of the expensive retrained model. This makes it useful for analysis of the unlearned model after deployment as well. We present results of experiments conducted for random subset forgetting and class forgetting on various deep networks and across different application domains.~Source code is at: https://github.com/vikram2000b/bad-teaching-unlearning
Introduction
Machine unlearning addresses the need to remove user information from trained models under data-deletion requirements, without retraining from scratch. The paper proposes a teacher-student approach with competent and incompetent teachers, a retrained-model-free metric, and broad evaluation across forgetting settings, networks, and domains.
- Motivation: Data-protection regulations support deletion requests and withdrawal of consent for data used in ML models.
- Motivation: Machine unlearning removes information derived from requested data, avoiding either full retraining or retaining that information in model parameters.
- Contributions: The method avoids constraints on the training procedure, trains no extra assisting models, and is evaluated across CNNs, vision transformers, LSTMs, and multimedia domains.
- Contributions: The proposed teacher-student framework selectively transfers knowledge from competent and incompetent teachers to support class-level and random-subset forgetting.
- Contributions: The paper introduces zero retrain forgetting (ZRF), a retrained-model-free metric for evaluating unlearning and its generalization on forget data.
Related Work
Prior deep-network unlearning methods use specialized training procedures, auxiliary models, or approximations. These requirements can increase computational costs and reduce practicality for real-world applications.
- Earlier deep unlearning work used scrubbing, neural tangent kernel approximations, or mixed-linear auxiliary models to modify network weights.
- Several existing methods require SGD-based optimization, additional models, or fine-tuning specific to different deep networks.
- These approaches are described as having high computational costs, training-process constraints, and limitations from approximation methods.
Preliminaries
The paper partitions the training data into forget and retain sets, assigns unlearning labels, and defines competent, incompetent, and student models for the proposed framework.
- Data and labels: The complete dataset D_c is partitioned into mutually exclusive forget and retain sets, D_f and D_r.
- Data and labels: Unlearning labels l_u mark forget samples with 1 and retain samples with 0.
- Teacher-student setup: The competent teacher is the fully trained original model, while the incompetent teacher is randomly initialized and the student starts with the original model’s parameters.
Proposed Method
The method uses competent and incompetent teachers to train a student that selectively removes forget-set information while preserving retain-set knowledge. It also introduces ZRF, a retrained-model-free metric whose interpretation depends on prediction randomness and expected generalization.
- Teacher-student unlearning: KL-Divergence terms define the teacher-student objectives for the incompetent teacher and competent teacher.The equations separately define divergence between the incompetent teacher and student, and between the competent teacher and student.
- Teacher-student unlearning: The student is trained with competent and incompetent teachers to remove information about forget data while retaining information about the retain set.The method transfers bad knowledge from the incompetent teacher for forget samples and accurate knowledge from the competent teacher for retain samples.
- Zero Retrain Forgetting: The ZRF metric evaluates unlearning without depending on an expensive retrained model by comparing unlearned-model predictions with an incompetent teacher.It uses Jensen–Shannon divergence to measure prediction randomness on the forget set.
- Zero Retrain Forgetting: ZRF scores range from 0 to 1, approaching 1 when forget-set predictions are completely random and 0 when they show a specific pattern.The metric compares the unlearned model’s forget-set output distribution with that of a randomly initialized incompetent teacher.
- Zero Retrain Forgetting: An ideal ZRF value depends on the model, dataset, and forget set, with test-set ZRF proposed as a practical proxy when retraining is unavailable.The test set is treated as data the model has never learned specifically and therefore as approximately perfectly unlearned.
Experiments
The experiments evaluate unlearning across datasets, architectures, forgetting settings, and metrics, comparing the proposed teacher-student method with retraining and Amnesiac learning. Results indicate close agreement with retrained models, reduced information exposure, and substantial runtime advantages.
- Experimental Setup: Experiments cover image classification, epileptic seizure recognition, and human activity recognition using CNNs, vision transformers, DNNs, and LSTMs.Datasets include CIFAR10, CIFAR100, epileptic seizure recognition, and activity recognition.
- Evaluation Measures: The evaluation measures forget and retain accuracy, membership inference, activation distance, JS-Divergence, and ZRF score.ZRF removes dependence on an available retrained model, while lower activation distance and JS-Divergence indicate closer agreement with retraining.
- Sample Unlearning: The proposed method achieves performance close to the retrained model when forgetting subclasses from CIFARSuper20.Experiments use 30% of retain data and one unlearning epoch, with results reported for ResNet18 and Vision Transformer models.
- Sample Unlearning: Membership inference probability on the forgotten Rocket class drops from 0.982 to 0.002, while JS-Divergence is 0.04 and ZRF rises from 0.87 to 0.99.The reported forget-set accuracy is almost the same as the retrained model.
- Comparison with Amnesiac Learning: For epileptic seizure forgetting, the proposed method reaches 74% forget-set accuracy versus 70% for retraining and 30% for Amnesiac learning.Activation distance is 0.47 for the proposed method and 0.77 for Amnesiac learning; the proposed method uses 30% of retain data and is approximately twice as fast.
- Efficiency Analysis: Random-weight unlearning is approximately 70× faster than retraining and more than 2× faster than Amnesiac learning, while a proxy-model setup is about 20× faster than retraining.Cheaper incompetent teachers can further accelerate unlearning, though partially trained teachers trade computational cost against closeness to the retrained model.
Conclusion
The paper presents a general teacher-student framework for machine unlearning and supports multiple forgetting settings across application domains and network modalities. It also introduces ZRF, which evaluates unlearning without requiring an expensive retrained model, while identifying improved efficiency and privacy guarantees as future work.
- The teacher-student framework supports single-class, multiple-class, subclass, and random-sample forgetting.
- Experiments evaluate the method across multiple application domains and network modalities.
- ZRF evaluates unlearning without requiring a retrained model for reference.The metric is intended for settings where retrained models are unavailable or expensive to obtain.
- Future work could improve the intersection of efficiency and privacy guarantees through better evaluation measures or new unlearning techniques.
Additional Ablation Studies
Additional experiments show that unlearning performance depends on retain-data volume, epochs, and learning rate. The selected settings can trade efficiency against similarity to the retrained model, with one configuration achieving JS-Divergence 0.03.
- Effect of Different Number of Epochs: More retain data and more unlearning epochs increase the ZRF score in the activity-recognition ablation.With 10% retain data, ZRF rises from 0.62 after 1 epoch to 0.75 after 5 epochs.
- Effect of Different Learning Rates: Changing the learning rate changes the amount of randomization in the updated model.The experiments fix unlearning at 2 epochs with 30% retain data.
- Effect of Different Learning Rates: 0.96 to 0.59: the ZRF score changes across learning rates 0.1 to 0.0001, respectively.
- JS-Divergence 0.03 is obtained with 30% retain data, 2 epochs, and a learning rate of 0.001.The authors identify this as the setting closest to the retrained model.
- Sequential requests remain close to retrained-model performance on both retain and forget sets.Figure 3 covers requests for Rocket, Mushroom, and Lamp subclasses.
Sequential Unlearning
The method remains robust under repeated unlearning requests. Forgetting Rocket, Mushrooms, and Lamp sequentially causes hardly any degradation in performance.
- Hardly any performance degradation is observed after sequentially forgetting Rocket, Mushrooms, and Lamp.The experiment simulates repeated real-life unlearning requests.
Partially Retrained Model as a Competent Teacher: Additional Experiments
The paper explores partially trained models as incompetent teachers for class unlearning. A teacher trained for 2 epochs on 50% of retain data is used in the reported experiment.
- Partially trained models on subsets of retain data can serve as incompetent teachers.Smaller models trained on small retain-data subsets are also proposed as alternatives.
- The reported partially trained teacher uses 2 epochs and 50% of the retain data.
Efficiency Analysis: More Details
The proposed method substantially reduces unlearning runtime compared with retraining and Amnesiac learning, while supporting broader forgetting modes than the fastest method, UNSIR.
- The proposed method takes substantially less time than both retraining and Amnesiac learning for class unlearning in ResNet18 on CIFAR10.
- The comparison evaluates total runtime across the retrained model, existing methods, and the proposed method.
- UNSIR is faster, but it supports only class-level unlearning, whereas the proposed method supports both sample-level and class-level unlearning.