Structural Fault Tolerance in Heterogeneous Financial AI Systems under Adversarial Signal Injection

Sammanfattning

Introduction: Financial AI systems increasingly combine large language models (LLMs), reinforcement learning (RL), and rule-based components within shared decision pipelines. Existing robustness mechanisms target individual models during training and do not address cross-modal failure propagation at inference time. Research Question: How can structural fault tolerance be designed and evaluated in heterogeneous financial AI systems to improve robustness under adversarial signal injection? Method: Following aDesign Science Research framework, this thesis evaluates a multi-agent trading architecture on Apple (AAPL) daily data and financial news. The system comprises an LLM-based Analyst (textual channel), an RL-based Executor (numerical channel), a rule-basedGuardian enforcing risk constraints, and a Consistency Gate (C-Gate) that arbitrates between the two reasoning channels at inference time. Three adversarial attack types at five intensities each target the semantic and numeric inputs. Results: The architecture outperforms a buy-and-hold benchmark on risk adjusted return while maintaining lower and more stable maximum drawdown (MaxDD) under adversarial conditions. Channel independence holds empirically: attacks on one channel do not degrade the other. Compared with an ablated configuration without the C-Gate, the full system achieves lower MaxDD, although no statistically significant Sharpe improvement is observed over the 127-day test period. Discussion: Architectural separation and inference-time arbitration improve robustness without additional model hardening. The C-Gate’s contribution is risk containment rather than return enhancement: it limits single-channel failure propagation and stabilises worst-case behaviour. The study is limited to a single asset and market regime, and the observed improvements remain statistically underpowered. Future work should evaluate the architecture across multiple assets, market conditions, and adaptive threat models.