RobustFTS: Defending Financial Time-Series Models Against Adversarial Manipulation

Deep learning models for financial time-series prediction are increasingly deployed in trading systems, yet their adversarial vulnerability remains poorly understood. We present RobustFTS, a defense framework combining: (i) volatility-scaled temporal randomization that injects noise proportional to local market volatility; (ii) financially-constrained adversarial training that restricts perturbations to market-realistic bounds; and (iii) a novel ensemble disagreement detector that exploits architecture-specific adversarial failures to flag attacks. RobustFTS achieves 73.2% robust prediction rate (60.8 percentage point improvement over the 12.4% baseline) with only 2% clean RMSE degradation, and detects attacks with 89.4% TPR at 5% FPR. We also find that foundation models (Lag-Llama, TimeGPT) are 18-21 percentage points more vulnerable than task-specific architectures. We release a robustness benchmark evaluating 6 architectures against 3 attack families. Scope: Evaluation focuses on intraday/next-day horizons; longer horizons and adaptive attacks require further study.