Relevance
2/10
The paper focuses on credit scoring and P2P lending rather than trading strategies or market prediction. However, the adversarial robustness methodology and multi-attack evaluation framework could be relevant to quantitative trading in terms of model robustness under adversarial market conditions, input manipulation in algorithmic trading systems, and risk management. The findings about cross-attack generalisation and mixed training strategies have methodological transferability to any ML-based financial decision system.
Implementation Complexity
6/10
The study involves three model families, four adversarial attack types plus a mixed regime, a full train-test grid (6 training configs × 5 test configs per model), stratified 5-fold cross-validation, and a detailed preprocessing pipeline. The adversarial attack implementations (FGSM, PGD, DeepFool, S&P) with feature masking for mutable-only perturbations add complexity. However, the individual components are well-established in the AML literature, and the code is publicly available. The main complexity lies in the systematic grid evaluation and ensuring structurally valid perturbations for categorical features.
Reproducibility
4/5
The paper provides a GitHub repository with full experimental code and configuration files. Uses a publicly available Lending Club dataset (Kaggle) and a Prosper dataset for external replication. Stratified 5-fold cross-validation is employed. Preprocessing pipeline is described in detail. Hyperparameters are reported in tables. Attack parameters are specified. However, the dataset is a curated subset (~400,000 observations) rather than the full dataset, and hyperparameter selection was done via manual pilot tuning rather than exhaustive search.