Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adversarial Robustness in Financial Machine Learning: Defenses, Economic Impact, and Governance Evidence

Samruddhi Baviskar

TL;DR

The paper tackles adversarial robustness in tabular financial machine learning by introducing a dataset-agnostic pipeline that combines gradient-based attacks (FGSM, PGD), domain-aware perturbation constraints, and both strong and lightweight defenses. It extends traditional performance evaluation with calibration, economic risk metrics (EL, VaR, ES), distribution drift, fairness, and explanation stability, including an LLM-based Semantic Robustness Index for governance signaling. Empirical results show that small perturbations degrade discrimination, calibration, tail risk, and explanation stability; PGD adversarial training improves resilience but does not fully close the robustness gap. The framework outputs governance-ready artifacts (bootstrap intervals, threshold-cost tables, economic confusion matrices) to support model risk management and regulatory reviews, representing a practical step toward robust, accountable financial ML systems.

Abstract

We evaluate adversarial robustness in tabular machine learning models used in financial decision making. Using credit scoring and fraud detection data, we apply gradient based attacks and measure impacts on discrimination, calibration, and financial risk metrics. Results show notable performance degradation under small perturbations and partial recovery through adversarial training.

Adversarial Robustness in Financial Machine Learning: Defenses, Economic Impact, and Governance Evidence

TL;DR

The paper tackles adversarial robustness in tabular financial machine learning by introducing a dataset-agnostic pipeline that combines gradient-based attacks (FGSM, PGD), domain-aware perturbation constraints, and both strong and lightweight defenses. It extends traditional performance evaluation with calibration, economic risk metrics (EL, VaR, ES), distribution drift, fairness, and explanation stability, including an LLM-based Semantic Robustness Index for governance signaling. Empirical results show that small perturbations degrade discrimination, calibration, tail risk, and explanation stability; PGD adversarial training improves resilience but does not fully close the robustness gap. The framework outputs governance-ready artifacts (bootstrap intervals, threshold-cost tables, economic confusion matrices) to support model risk management and regulatory reviews, representing a practical step toward robust, accountable financial ML systems.

Abstract

We evaluate adversarial robustness in tabular machine learning models used in financial decision making. Using credit scoring and fraud detection data, we apply gradient based attacks and measure impacts on discrimination, calibration, and financial risk metrics. Results show notable performance degradation under small perturbations and partial recovery through adversarial training.

Paper Structure

This paper contains 47 sections, 19 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Adversarial Robustness in Machine Learning
  4. Adversarial Attacks on Tabular and Structured Data
  5. Financial ML and Risk-Sensitive Robustness
  6. Calibration, Uncertainty, and Distribution Shift
  7. Explanation Robustness Under Attack
  8. Robust Defenses and Practical Mitigation
  9. Threat Model
  10. Perturbation Model
  11. Attacker Capabilities
  12. Evaluation Objectives
  13. Illustrative Diagram of the Threat Model
  14. Data and Pre-Processing
  15. Pre-Processing Pipeline
  16. ...and 32 more sections

Figures (2)

  • Figure 1: Schematic of adversarial example generation in tabular ML. Perturbations are constrained by financial plausibility and projected into valid feature domains before gradient-based optimization.
  • Figure 2: Overall architecture of the proposed adversarial robustness pipeline for financial machine learning. The framework integrates domain-aware adversarial attacks, robust training, economic risk evaluation, and explanation stability analysis.