Paper: arXiv 2308.11406

Authors: Alexey Zaytsev, Maria Kovaleva, Alex Natekin, Evgeni Vorsin, Valerii Smirnov, Georgii Smirnov, Oleg Sidorshin, Alexander Senin, Alexander Dudin, Dmitry Berestnev

Abstract

Banks routinely use neural networks to make decisions. While these models offer higher accuracy, they are susceptible to adversarial attacks, a risk often overlooked in the context of event sequences, particularly sequences of financial transactions, as most works consider computer vision and NLP modalities. We propose a thorough approach to studying these risks: a novel type of competition that allows a realistic and detailed investigation of problems in financial transaction data. The participants directly oppose each other, proposing attacks and defenses – so they are examined in close-to-real-life conditions. The paper outlines our unique competition structure with direct opposition of participants, presents results for several different top submissions, and analyzes the competition results. We also introduce a new open dataset featuring financial transactions with credit default labels, enhancing the scope for practical research and development.

Complexity vs Empirical Score

  • Math Complexity: 2.5/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper’s core contribution is a novel competition framework and an open dataset for testing model robustness, focusing on implementation and data rather than advanced mathematical theory. Empirical rigor is high due to the introduction of a new dataset, detailed competition results, and analysis of participant submissions under realistic adversarial conditions.

Research Flowchart

  flowchart TD
  A["Research Goal: How to improve robustness of<br>financial transaction models against adversarial attacks?"] --> B["Methodology: Direct Opposition Competition"]
  
  B --> C["Data Input:<br>Open Financial Transaction Dataset<br>(with credit default labels)"]
  
  C --> D["Process: Attackers & Defenders<br>compete in real-time"]
  
  D --> E["Computational Process:<br>Neural Network Training &<br>Adversarial Evaluation"]
  
  E --> F["Outcome 1: Novel Competition Structure<br>validated for financial ML"]
  
  E --> G["Outcome 2: Performance Analysis<br>of top attack/defense submissions"]
  
  E --> H["Outcome 3: Enhanced Robustness<br>framework for banking applications"]
  
  F --> I["Key Finding:<br>Direct opposition reveals<br>real-world vulnerabilities<br>in event sequence models"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a novel competition framework to evaluate robustness of financial transaction models and releases a new open dataset of bank transactions with credit default labels.

Source sentence
a novel type of competition that allows a realistic and detailed investigation of problems in financial transaction data.

Assumes 4

  • The competition uses a static model environment in the pre‑tournament phase. model
    Source sentence
    The pre‑tournament phase allows for detailed study of a static model environment
  • Each customer sequence contains at least 300 transactions. data
    Source sentence
    Each customer within the dataset has at least 300 associated transactions in a sequence.
  • Transaction amounts are anonymized with normalization and small noise. data
    Source sentence
    all transaction amounts within the dataset have been anonymized with normalization and small noise.
  • The credit default positive label rate is 0.04. data
    Source sentence
    the share of the positive label rate of defaults in data is 0.04.

Data

  • Default other public
    Source sentence
    For this competition, we present a new open dataset of bank transactions named Default, which is described below.
  • Churn other public
    Source sentence
    Also, we applied a previously existing dataset named Churn for additional testing of the best solutions.

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • The highest defense score achieved was 0.72, exceeding the quality of the initial model, but defenses significantly degrade against stronger attacks. performance
    Source sentence
    while weaker attacks lead to minor effects, with the highest score of 0.72 being higher than the quality of the initial model, the defenses significantly degrade when put against stronger attacks.
  • During the black‑box tournament stage, ROC AUC decreased on average by about 0.05 across attack‑defense pairs, with some attacks and defenses showing almost zero decrease. performance
    Source sentence
    we still see defenses and attacks with almost zero decrease in ROC AUC and about a 0.05 decrease of ROC AUC on average
  • Models using only features not susceptible to attack (Boosting clean) achieve competitive ROC AUC on clean data, and gradient‑boosted models distilled from neural networks improve performance. predictive-accuracy
    Source sentence
    For the sake of comparison, we also present scores for Boosting clean model that uses only features not susceptible to the attack

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 11 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.