Papers
Topics
Authors
Recent
Search
2000 character limit reached

When AI Meets Wall Street: A Survey on Trustworthy AI in Fintech

Published 28 May 2026 in cs.CR | (2605.30650v1)

Abstract: Artificial intelligence is now embedded as a primary decision engine in continuously operated financial AI pipelines spanning training and updating, deployment and inference, and operation with monitoring and feedback. The automation and scale that make these pipelines effective also create novel attack surfaces, where small algorithmic perturbations can amplify into persistent, system-level financial harm. Existing surveys, however, either treat AI as a defensive tool or analyse adversarial machine learning in a domain-agnostic manner, abstracting away finance-specific constraints such as accounting plausibility, non-IID federated data, continuous retraining, and automation-amplified downstream effects. We address this gap with a unified, lifecycle-centric and mechanism-driven framework. We partition financial AI into three lifecycle stages: training and updating, deployment and inference, and operation, monitoring, and feedback. We further propose the Financial AI Security and Robustness Taxonomy, organising seventeen attack subtypes across data and model poisoning, adversarial attacks on decision boundaries, prompt injection in LLM-mediated workflows, and deepfake-driven subversion of KYC verification layers. For each subtype, we analyse algorithmic strategy, feasibility constraints, stealth and persistence, and downstream financial consequences. Finally, we identify open challenges and outline a research agenda toward lifecycle-aware stress testing and finance-relevant robustness benchmarks.

Summary

  • The paper develops a lifecycle-based taxonomy of 17 attack subtypes across training, deployment, and operational phases, linking mechanisms such as poisoning, evasion, prompt injection, and deepfakes to financial and systemic consequences.
  • The paper finds that small, feasible manipulations can cause severe damage, including label poisoning that cuts fraud recall from 88% to 47%, transfer attacks that reduce recall from 92% to 56%, and backdoors exceeding 90% attack success rates.
  • The paper argues that single-metric defenses are routinely bypassed and recommends lifecycle-aware stress testing that combines cross-round auditing, replay checks, source-integrity verification, stateful monitoring, and finance-specific robustness benchmarks.

This survey develops a lifecycle-centric and mechanism-driven framework for analyzing security and robustness threats in continuously operated financial AI pipelines (2605.30650). Its central premise is that financial AI systems are no longer isolated models but automation-amplified socio-technical pipelines—spanning training and updating, deployment and inference, and operation with monitoring and feedback—in which small algorithmic perturbations can propagate through long decision chains and persist via feedback loops. The authors argue that existing literature exhibits a structural mismatch: finance-facing security research treats AI primarily as a defensive tool, while adversarial machine learning (AML) taxonomies remain domain-agnostic and abstract away finance-specific constraints such as accounting plausibility, non-IID federated data, continuous retraining, and asymmetric error costs.

Methodology

The survey follows a structured review protocol inspired by PRISMA-style screening over Google Scholar, ACM Digital Library, IEEE Xplore, SpringerLink, ScienceDirect, arXiv, and major security/ML venues, augmented by backward and forward snowballing. Inclusion requires an explicit threat or failure model, a concrete attack or defense mechanism applicable to financial AI tasks, and supporting evaluation evidence; purely policy-oriented work and studies treating AI solely as a defensive tool are excluded. For each retained paper, the authors extract lifecycle stage, attack entry point, exploited mechanism, threat model and feasibility constraints, complexity profile, and downstream financial consequence. The authors concede three limitations of this procedure: LLM-agent workflow attacks evolve faster than peer review, proprietary financial datasets restrict reproducibility, and many defenses are validated only on simplified benchmarks rather than real operating conditions.

The Financial AI Security and Robustness Taxonomy

The organizing contribution is a taxonomy of seventeen attack subtypes (S1–S17) mapped onto three lifecycle phases:

Phase Attack class Subtypes
Training & updating Data poisoning Label-level (S1), feature-level (S2), instance-level (S3)
Training & updating Model poisoning Global gradient (S4), fragment/coordinate-level (S5), backdoor/trigger-based (S6)
Deployment & inference Adversarial attacks on decision boundaries Input manipulation (S7), single-step (S8), optimisation-based (S9), transfer-based (S10)
Operation & monitoring Prompt injection in LLM workflows Character/symbol (S11), word/phrase (S12), sentence/instruction (S13), chain/indirect (S14)
Operation & monitoring Deepfake subversion of KYC Latent identity synthesis (S15), feature/attribute transfer (S16), adversarial anti-forensic generation (S17)

For each subtype the survey analyzes algorithmic strategy, computational complexity, stealth and persistence mechanisms, and downstream financial consequences, assigning a systemic risk priority to each.

Threats in the training and updating phase

The survey's treatment of data poisoning contrasts optimization-precision against stochastic scalability. Gallagher et al.'s profit-based targeted label injection for time-series forecasting flips labels on the least profitable samples at O(Nlog⁡N)O(N\log N) cost, and flipping only 10% of samples increases trading losses by more than 130%. Kure et al. show that random label flipping at linear O(N)O(N) complexity achieves comparable damage without any oracle: on insurance fraud detection, flipping just 5% of labels collapses fraud recall from 88% to 47%. At the feature level, Alfeld et al.'s convex-programming poisoning of autoregressive models shows that recursive prediction amplifies bounded early-step perturbations into systematic forecast bias under budgets as low as 10%, while Sonic uses incremental density proxies and genetic search to disrupt clustering structure orders of magnitude faster than direct attacks—a mechanism that transfers to AML community detection. Instance-level attacks carry the highest systemic priority: model hijacking reaches roughly 99% success with only 0.5% clean-accuracy loss, trigger-based face backdoors approach 100% success at negligible inference overhead, and black-box incremental poisoning raises misclassification from 0% to 30% over fifteen retraining rounds, exploiting history dependence in online fine-tuning.

For federated model poisoning, the survey identifies a structural blind spot—"data availability without data visibility"—in which servers cannot inspect local data and must rely on statistical aggregation. PoisonedFL degrades models to random-guess capability while bypassing Trimmed Mean and FLTrust via cross-round directional consistency; ScaleSign achieves 98.23% attack success against SignGuard by mimicking benign sign statistics; FMPA concentrates manipulation on k≪dk \ll d high-leverage coordinates at O(k)O(k) per-round cost; and Sine demonstrates that cosine-similarity-to-anchor checks alone are bypassable, with a paired defense reducing residual impact to roughly 2–4%. Backdoor attacks bifurcate by topology: in vertical FL, clean-label universal triggers exceed 90% ASR on LendingClub while preserving AUC by inferring pseudo-labels from gradient norms; in horizontal FL, density-adaptive distributed triggers reach 91.68% ASR while outperforming DBA baseline accuracy by over 8%. The implication drawn is that single-metric defenses (cosine similarity, sign distributions, norm bounds) are each demonstrably bypassable, so consortium auditing must combine cross-round sign-distribution drift, layer-wise residual energy, and periodic replay checks.

Threats in the deployment and inference phase

The deployment-phase analysis covers four adversarial attack families. Input perturbation attacks include the Single Value Attack, which perturbs only the most gradient-influential timestep within a sliding window; a GAN-based Feature Selector Network that reduces black-box query complexity by approximately 75% and cuts adversarial training success rates by roughly 40% on IEEE-CIS fraud data; and the Maximum Violated Multi-Objective attack on financial reporting, which jointly inflates EPS and reduces the Beneish M-score in about 50% of trials under accounting-consistency constraints—the first formalization of evasion against regression models bound by accounting identities. Single-step FGSM-style attacks exploit local linearity: applied to stock forecasting with volatility-calibrated magnitudes, they produce cumulative simulated losses of approximately $200k, and tokenized transaction attacks modify only one or two records to achieve black-box success rates above 0.8.

Optimisation-based attacks extend PGD to order-book manipulation under market microstructure constraints (integer order sizes, capital budgets); notably, MLPs prove less robust than linear models despite higher clean accuracy, and LSTMs exhibit gradient masking yet remain highly vulnerable to transfer attacks. Transfer-based results carry the strongest defensive implication: adversarial samples crafted on a logistic-regression surrogate transfer with 94% efficacy to a random forest target in credit card fraud detection, dropping recall from 92% to 56%—undermining heterogeneous ensembling as a defense—and DRL trading agents exhibit high transferability even from imitated surrogates, with net-worth impact often invisible in the reward metric used for human oversight. The survey concludes that architectural secrecy provides no protection and that universal/cross-architecture transfer attacks represent the highest systemic risk in this phase.

Threats in operation, monitoring, and feedback

The operational phase covers two vectors exploiting "semantic trust." Prompt injection is organized across four granularity levels. Character-level encoding attacks (zero-width characters, Unicode tag smuggling, emoji smuggling) achieve reported bypass rates up to 100% against guardrail detectors at negligible cost. Word/phrase-level attacks bifurcate between computationally heavy latent-space semantic optimization for event-driven stock prediction and O(1)O(1) homoglyph headline manipulation that can misroute signals in latency-sensitive algorithmic trading. Instruction-level attacks progress from template encapsulation (95%+ ASR) to gradient-optimized system-prompt leaking, implying attackers can systematically extract hidden risk-control logic. Chain/indirect injection is assigned the highest priority because payloads planted in RAG documents or API outputs hijack tool-using agents without touching the user prompt; memory-augmented black-box attacks reach approximately 71.9% average ASR on commercial agents, transferring directly to financial report summarization and KYC document parsing.

Deepfake attacks target the KYC verification layer through three subtypes. Latent identity synthesis includes Arc2Face's identity-controlled diffusion generation—which produces photorealistic faces optimized against the same ArcFace embeddings used in banking recognition—and master-face evolution that falsely matches significant population fractions, threatening 1:N deduplication. Feature-level transfer is rated most severe: DiffSwap preserves 3D pose and micro-expressions to neutralize active liveness detection, and LiveBugger demonstrates that leading commercial facial liveness APIs prioritize action correctness over source authenticity, enabling high bypass rates in remote KYC and account recovery. Adversarial anti-forensic generation spans physical optical projection attacks on structured-light 3D sensors and query-efficient decision-based evasion suited to black-box verification APIs. The unifying consequence is that the "what you see is who is there" assumption underlying remote verification no longer holds, forcing governance toward source-integrity verification, virtual-camera detection, rPPG analysis, and stateful query monitoring.

Positioning relative to prior surveys

The survey differentiates itself along four axes from adjacent literatures: FinTech cybersecurity surveys treat AI as a defensive tool; broad AI-in-finance surveys treat security as one component among adoption and regulation topics; financial XAI surveys address interpretability rather than robustness (a model may be explainable yet fragile); and general AML meta-surveys are necessarily domain-agnostic, omitting finance-specific constraints such as real-time execution, asymmetric error costs, auditability obligations, and correlated-failure systemic risk. The distinctive design choice is connecting mechanism-level vulnerabilities to concrete lifecycle stages and downstream business, regulatory, and systemic outcomes.

Limitations and open questions

Beyond the methodological limitations noted above, the survey identifies three substantive open problems. First, persistent manipulation across updates remains insufficiently understood in continuous-retraining, federated-aggregation, and long-horizon feedback settings—for example, whether cheap stochastic poisoning can match targeted sorting-based attacks, and how backdoors survive after malicious participants exit a federation. Second, the robustness of monitoring signals themselves is under-examined: adaptive adversaries can bypass, delay, or corrupt anomaly detection, and several reviewed attacks explicitly defeat perplexity filters and statistical aggregators. Third, emerging threats in AI-mediated workflows—retrieval-chain hijacking, defense-aware indirect injection, deepfake-enabled identity subversion—extend the attack surface beyond model parameters into workflow logic and evidentiary interfaces, and current defenses are validated mostly on simplified benchmarks rather than production financial conditions. The paper also leaves open whether finance-relevant robustness benchmarks and lifecycle-aware stress-testing protocols can be standardized given proprietary data constraints.

Conclusion

This survey consolidates a fragmented literature into a unified taxonomy linking seventeen attack subtypes to three stages of continuously operated financial AI pipelines, with explicit attention to complexity profiles, stealth mechanisms, and downstream financial consequences. Its recurring analytical findings—that small perturbations amplify through recursion and feedback, that single-metric defenses are systematically bypassable, and that the highest-priority risks arise from function-rewriting attacks (backdoors, chain injection, liveness-neutralizing face swaps) rather than indiscriminate corruption—provide a coherent basis for prioritizing defense investment. The proposed agenda of lifecycle-aware stress testing and finance-relevant robustness benchmarks defines the immediate research gap the survey leaves to subsequent work.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.