Fast Unlearning Techniques for Neural Network Prediction and Classification Algorithms

Enforcing the forgetting (or unlearning) of specific portions of a neural network's (NN) memory is necessary for both security and ethical considerations. Moreover, targeted forgetting can enhance the network's performance by eliminating the learning derived from compromised data. Conventional approaches, such as retraining the model from scratch, are associated with significant time and energy costs, while alternative methods can degrade the neural network's performance in terms of complexity and predictive accuracy. To address these challenges, this work introduces two novel approaches that mitigate the drawbacks associated with traditional unlearning methods. Both approaches leverage the Long Short-Term Memory (LSTM) network, chosen for its superior learning quality and efficient processing time—crucial factors in such algorithms. The first approach involves modifying the LSTM architecture by coarse tune the weight matrix of the forget gate to facilitate controlled forgetting. The second approach follows the idea of intruders by inducing poluuted data to the system, and the idea is based on learning the model through corrupted/bad data, which causes the LSTM memory to be changed at specific time intervals that correspond to the data that should be forgotten. We present a performance evaluation of the second approach, with a future study planned for the performance assessment of the first approach. The evaluation demonstrates the superiority of the proposed solutions in terms of quality of service and implementation simplicity.