Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Saptarshi Sengupta

Second Advisor

Amith Kamath Belman

Third Advisor

Maryam Khazaei

Keywords

Moving Target Defense, Time-Series Forecasting, White-box attacks, Game-Theoretic Scheduling, Bayesian Attacker Modeling, Task-Aware Optimization

Abstract

Deep learning models have been extensively used in time-series forecasting tasks, such as electricity demand prediction and weather forecasting. However, these models are vulnerable to adversarial attacks, in which minor perturbations in the input data may cause significant degradation in the forecasting performance. Existing defense methods enhance robustness, but most of them are static and fail to adapt to different attack scenarios. This work proposes a task-aware, game-theoretic scheduling framework for adversarially robust time-series forecasting. The approach is based on a Moving Target Defense (MTD) paradigm, where a pool of student models is generated by structured perturbations of a base model. Instead of uniform random switching to select models, the proposed framework uses a Bayesian game-theoretic scheduler that assigns probabilities to models according to attacker priors and task-specific payoff functions, including horizon-aware and channel-aware objectives. The proposed method is evaluated on the Electricity Load and Jena Climate datasets under multiple adversarial attack settings. Performance is measured in terms of RMSE, hmax, cmax and hwmean. Experimental results show that the scheduler is consistently better than uniform random switching and achieves similar performance to the best single model across different attacker priors.

Download

Available for download on Saturday, May 22, 2027

Share

COinS