Paper: arXiv 2407.14573

Authors: Orson Mengara

Abstract

Since the advent of generative artificial intelligence, every company and researcher has been rushing to develop their own generative models, whether commercial or not. Given the large number of users of these powerful new tools, there is currently no intrinsically verifiable way to explain from the ground up what happens when LLMs (large language models) learn. For example, those based on automatic speech recognition systems, which have to rely on huge and astronomical amounts of data collected from all over the web to produce fast and efficient results, In this article, we develop a backdoor attack called MarketBackFinal 2.0, based on acoustic data poisoning, MarketBackFinal 2.0 is mainly based on modern stock market models. In order to show the possible vulnerabilities of speech-based transformers that may rely on LLMs.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 2.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper introduces advanced financial mathematics (e.g., stochastic calculus, optimal transport, Bayesian optimization) but lacks any backtesting data, performance metrics, or implementation details, focusing instead on theoretical attack methodology.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Vulnerability of Speech LLMs"] --> B["Data: <br>Acoustic Datasets & Stock Market Data"]
  B --> C["Methodology:<br>Bayesian Optimization &<br>Acoustic Poisoning"]
  C --> D["Computation:<br>Backdoor Injection<br>MarketBackFinal 2.0"]
  D --> E["Outcome:<br>Identified Vulnerabilities<br>in Speech Transformers"]

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 backdoor attack called MarketBackFinal 2.0, based on acoustic data poisoning and modern stock market models.

Source sentence
For example, those based on automatic speech recognition systems, which have to rely on huge and astronomical amounts of data collected from all over the web to produce fast and efficient results, In this article, we develop a backdoor attack called MarketBackFinal 2.0, based on acoustic data poisoning, MarketBackFinal 2.0 is mainly based on modern stock market models.

Assumes 3

  • The model owner is the defender and the trusted third party is the attacker in the threat model. model
    Source sentence
    We assume that the model owner and the trusted third party play the roles of defender and attacker, respectively, in the threat model.
  • In a white‑box setting the adversary can access and modify the dataset, model parameters and structure. other
    Source sentence
    In a white-box environment, the adversary understands and controls the target dataset and model, including the ability to access and modify the dataset, as well as parameters and the structure of the model.
  • In a black‑box setting the adversary can only manipulate part of the training data and has no knowledge of model structure or parameters. other
    Source sentence
    However, in the stricter framework of the black-box, the attacker is only able to manipulate part of the training data but has no knowledge of the structure and parameters of the target model, such as weights, hyperparameters, configurations, etc.

Data

  • ESC-50 environmental dataset benchmark public
    Source sentence
    We use the ESC-50 environmental dataset [66], the dataset ESC-50 (environmental), is a labeled collection of 2,000 environmental sounds, which comprises five-second recordings sorted into 50 semantic categories (with 40 examples per category) that are further divided into five main groups: animals; natural soundscapes and water sounds; humans, non-vocal sounds; indoor/domestic sounds; and outdoor/urban sounds.

Costs

Not applicable The paper reports no trading or execution result.

Finds 2

  • The backdoor attack achieves near‑perfect attack success rate (100%) while maintaining high benign accuracy (≈94–99%) across multiple pre‑trained speech models. performance figure from elsewhere in the paper
    Source sentence
    Hugging Face Models Benign Accuracy (BA) Attack Success Rate (ASR) wav2vec 2.0 94.73% 100% whisper (OpenAI) 95.03% 100% HuBERT 95.21% 100% facebook/w2v-bert-2.0(Facebook) 98.96% 100% facebook/mms-1b-all 93.31% 100% Speech Encoder Decoder 96.12% 100% Data2vec 99.12% 100% 2 ESC-50 environmental dataset.
  • The authors state that MarketBackFinal 2.0 easily misleads almost all tested Hugging Face transformer models. other figure from elsewhere in the paper
    Source sentence
    We can see that our backdoor 18Hugging Face Transformers THE FINAL PREPRINT , JOURNAL NAME ——-, VOL. .., NO. ..., MONTH 2024 5 attack easily manages to mislead these models (readers are invited to test 19), other Hugging Face models; as far as we know, we’ve managed to fool almost all these models.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 6 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.

Cited References (125)

  1. Lifting the Heston model (2019)
  2. On simulation of rough Volterra stochastic volatility models (2021) — link
  3. Volatility models in practice: Rough, Path-dependent or Markovian? — → 2401.03345
  4. Volatility is rough (2018)
  5. Actuarial applications of natural language processing using transformers: Case studies for using text features in an actuarial context (2024)
  6. A partial rough path space for rough volatility (2024)
  7. Rough volatility: fact or artefact? (2024)
  8. Multiscaling and rough volatility: An empirical investigation (2022)
and 117 more
  1. Multifactor approximation of rough volatility models (2019)
  2. Short-term at-the-money asymptotics under stochastic volatility models (2019)
  3. The characteristic function of rough Heston models (2019)
  4. The microstructural foundations of leverage effect and rough volatility (2018)
  5. Perfect hedging in rough Heston models (2018)
  6. Markovian structure of the Volterra Heston model (2019)
  7. Explainable bayesian optimization (2024) — link
  8. The unreasonable effectiveness of optimal transport in economics (2021) — link
  9. Functional central limit theorems for rough volatility (2024)
  10. Rough volatility (2023)
  11. New and robust drift approximations for the LIBOR market model (2008)
  12. How to keep text private? A systematic review of deep learning methods for privacy-preserving natural language processing (2023)
  13. Backdoor attacks and defenses in federated learning: Survey, challenges and future research directions (2024)
  14. A survey on XAI for beyond 5G security: technical aspects, use cases, challenges and research directions (2022) — link
  15. A survey on llm-gernerated text detection: Necessity, methods, and future directions (2023) — link
  16. Machine and deep learning for iot security and privacy: applications, challenges, and future directions (2022)
  17. Privacy-preserving machine learning: Methods, challenges and directions (2021) — link
  18. Gpt (generative pre-trained transformer)–a comprehensive review on enabling technologies, potential applications, emerging challenges, and future directions (2024)
  19. A survey of privacy attacks in machine learning (2023)
  20. Large language models for software engineering: A systematic literature review (2023) — link
  21. A survey on large language model (llm) security and privacy: The good, the bad, and the ugly (2024)
  22. Blockchain-based federated learning: A comprehensive survey (2021) — link
  23. A survey of safety and trustworthiness of large language models through the lens of verification and validation (2024)
  24. Sleeper agents: Training deceptive llms that persist through safety training (2024) — link
  25. From chatgpt to threatgpt: Impact of generative ai in cybersecurity and privacy (2023)
  26. A review of speech-centric trustworthy machine learning: Privacy, safety, and fairness (2023)
  27. Trustllm: Trustworthiness in large language models (2024) — link
  28. On the opportunities and risks of foundation models (2021) — link
  29. Deep hedging: learning to simulate equity option markets (2019) — link
  30. Ai in finance: challenges, techniques, and opportunities (2022)
  31. Sustainable ai: Environmental implications, challenges and opportunities (2022)
  32. Large language models and knowledge graphs: Opportunities and challenges (2023) — link
  33. AI and 6G into the metaverse: Fundamentals, challenges and future research trends (2024)
  34. Optimal static-dynamic hedges for exotic options under convex risk measures (2009)
  35. A survey on optimal transport for machine learning: Theory and applications (2021) — link
  36. Bayesian optimization of ESG financial investments — → 2303.01485
  37. A tutorial on Bayesian optimization (2018) — link
  38. The Art of Deception: Robust Backdoor Attack using Dynamic Stacking of Triggers (2024) — link
  39. The last Dance: Robust backdoor attack via diffusion models and bayesian approach (2024) — link
  40. Trading Devil: Robust backdoor attack via Stochastic investment models and Bayesian approach — → 2406.10719
  41. ESC: Dataset for environmental sound classification (2015)
  42. Robust speech recognition via large-scale weak supervision (2023)
  43. Seamless: Multilingual Expressive and Streaming Speech Translation (2023) — link
  44. Scaling speech technology to 1,000+ languages (2024)
  45. wav2vec 2.0: A framework for self-supervised learning of speech representations (2020)
  46. Hubert: Self-supervised speech representation learning by masked prediction of hidden units (2021)
  47. Data2vec: A general framework for self-supervised learning in speech, vision and language (2022)
  48. Wav2seq: Pre-training speech-to-text encoder-decoder models using pseudo languages (2023)
  49. Theory and practice of model risk management (2002)
  50. Certified Robustness to Data Poisoning in Gradient-Based Training (2024) — link
  51. Generalization Bound and New Algorithm for Clean-Label Backdoor Attack (2024) — link
  52. A comprehensive overview of backdoor attacks in large language models within communication networks (2024)
  53. Can you hear it? backdoor attacks via ultrasonic triggers (2022)
  54. Audio-domain position-independent backdoor attack via unnoticeable triggers (2022)
  55. Bayesian inference based on stationary fokker-planck sampling (2010)
  56. A comprehensive survey on applications of transformers for deep learning tasks (2023)
  57. Hugging face (2022)
  58. Opportunistic Backdoor Attacks: Exploring Human-imperceptible Vulnerabilities on Speech Recognition Systems (2022)
  59. Badgpt: Exploring security vulnerabilities of chatgpt via backdoor attacks to instructgpt (2023) — link
  60. Badpre: Task-agnostic backdoor attacks to pre-trained nlp foundation models (2021) — link
  61. Hidden killer: Invisible textual backdoor attacks with syntactic trigger (2021) — link
  62. Attacks in adversarial machine learning: A systematic survey from the life-cycle perspective (2023) — link
  63. A Survey of Backdoor Attacks and Defenses on Large Language Models: Implications for Security Measures (2024) — link
  64. Large language models in finance: A survey — → 2311.10723
  65. A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges — → 2406.11903
  66. A survey of large language models in finance (finllms) (2024) — link
  67. Security and privacy challenges of large language models: A survey (2024) — link
  68. A survey on deep learning for financial risk prediction (2021)
  69. Delta hedging in financial engineering: Towards a model-free approach (2010)
  70. Preliminary remarks on option pricing and dynamic hedging (2012)
  71. Taking the human out of the loop: A review of Bayesian optimization (2015)
  72. Video Diffusion Models: A Survey (2024) — link
  73. A survey on video diffusion models (2023) — link
  74. Text-to-image diffusion models in generative ai: A survey (2023) — link
  75. A survey on generative diffusion models (2024)
  76. Diffusion models: A comprehensive survey of methods and applications (2023)
  77. Girsanov’s Theorem (2016)
  78. Implicit Diffusion: Efficient Optimization through Stochastic Sampling (2024) — link
  79. Large language models are efficient learners of noise-robust speech recognition (2024) — link
  80. Detectors for safe and reliable llms: Implementations, uses, and limitations (2024) — link
  81. Universal vulnerabilities in large language models: Backdoor attacks for in-context learning (2024) — link
  82. Foundational challenges in assuring alignment and safety of large language models (2024) — link
  83. Training compute-optimal large language models (2022) — link
  84. The Shadow of Fraud: The Emerging Danger of AI-powered Social Engineering and its Possible Cure (2024) — link
  85. Comparison of Static Application Security Testing Tools and Large Language Models for Repo-level Vulnerability Detection (2024) — link
  86. Imposter. AI: Adversarial Attacks with Hidden Intentions towards Aligned Large Language Models (2024) — link
  87. Targeted Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs (2024) — link
  88. Operationalizing a Threat Model for Red-Teaming Large Language Models (LLMs) (2024) — link
  89. Deep reinforcement learning for automated stock trading: An ensemble strategy (2020)
  90. Financial statement analysis with large language models — → 2407.17866
  91. BaThe: Defense against the Jailbreak Attack in Multimodal Large Language Models by Treating Harmful Instruction as Backdoor Trigger (2024) — link
  92. Transferring Backdoors between Large Language Models by Knowledge Distillation (2024) — link
  93. BadMerging: Backdoor Attacks Against Model Merging (2024) — link
  94. A rough path perspective on renormalization (2019)
  95. Learning tensor networks with parameter dependence for Fourier-based option pricing — → 2405.00701
  96. Fast Fourier transform option pricing with stochastic interest rate, stochastic volatility and double jumps (2013)
  97. An FFT-network for L 'e vy option pricing (2011)
  98. Pricing Israeli Option with Time-changed Compensation by an FFT-Based High-order Multinomial Tree in L 'e vy Markets (2022)
  99. A Fourier transform method for spread option pricing (2010)
  100. Signature volatility models: pricing and hedging with Fourier (2024)
  101. Ruin probabilities in the Cram 'e r–Lundberg model with temporarily negative capital (2020)
  102. The Cram 'e r-Lundberg model with a fluctuating number of clients (2023)
  103. Ruin probability in the Cram 'e r–Lundberg model with risky investments (2011)
  104. Trading and arbitrage in cryptocurrency markets (2020)
  105. Critical probabilistic characteristics of the Cram 'e r model for primes and arithmetical properties (2024)
  106. Neural McKean-Vlasov Processes: Distributional Dependence in Diffusion Processes (2024)
  107. Pathwise McKean-Vlasov Theory with Additive Noise (2018) — link
  108. Control of McKean–Vlasov dynamics versus mean field games (2013)
  109. Learning high-dimensional McKean–Vlasov forward-backward stochastic differential equations with general distribution dependence (2024)
  110. McKean–Vlasov type stochastic differential equations arising from the random vortex method (2022)
  111. Large deviations and exit-times for reflected McKean–Vlasov equations with self-stabilising terms and superlinear drifts (2022)
  112. Maximum principle for optimal control of McKean-Vlasov FBSDEs with L 'e vy process via the differentiability with respect to probability law (2019)
  113. The rough Hawkes Heston stochastic volatility model (2024)
  114. Volatility options in rough volatility models (2020)
  115. On spatially irregular ordinary differential equations and a pathwise volatility modelling framework (2019) — link
  116. Affine rough models (2018) — link
  117. A geometric perspective on diffusion models (2023) — link

7 of 125 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.