TrojanWave: Exploiting Prompt Learning for Stealthy Backdoor Attacks on Large Audio-Language Models

Asif Hanif, Maha Tufail Agro, Fahad Shamshad, Karthik NandakumarEMNLP 2025

TL;DR

This paper introduces TrojanWave, the first method to inject stealthy backdoor attacks into large audio-language models (ALMs) by manipulating only their learnable prompts, not the frozen model weights. They also propose TrojanWave-Defense, a lightweight defense to neutralize these prompt-based backdoors.

What Problem Does This Paper Address?

Large audio-language models (ALMs) are increasingly used in critical applications like security and healthcare. While prompt learning offers an efficient way to adapt these models, it also opens a new attack vector: backdoor attacks embedded within prompts. Unlike traditional backdoor attacks that require modifying model weights during training, TrojanWave exploits the prompt-learning paradigm, where users adopt pre-trained, frozen ALMs with potentially malicious prompts. This makes the attack stealthier and more scalable, especially in "Prompt-as-a-Service" (PaaS) scenarios, where users might unknowingly download compromised prompts. The paper addresses the critical security vulnerability of ALMs in the context of prompt learning, which has been largely unexplored.

Methods

The core of TrojanWave is its novel attack strategy. Instead of retraining the entire model, it focuses on manipulating the learnable prompts used in prompt learning. The attack embeds imperceptible audio triggers in both the time and spectral domains of the audio waveform. These triggers are then associated with a specific, attacker-chosen target label by optimizing the learnable prompts during a few-shot training phase. This process effectively "hijacks" the model's behavior, causing it to misclassify inputs containing the trigger while performing normally on clean inputs.

To counter this threat, the authors propose TrojanWave-Defense. This is a lightweight, post-hoc defense mechanism that purifies compromised prompts. It works by fine-tuning the infected prompt on a clean few-shot dataset, using a combination of a standard cross-entropy loss (to maintain task performance) and a "context-repulsion" term. This repulsion term actively pushes the purified prompt away from the original, backdoor-infected prompt, thereby weakening the backdoor's association. Crucially, this defense does not require access to poisoned data and operates in a few-shot setting.

Datasets

The paper evaluates TrojanWave and its defense across 11 diverse audio classification datasets. These datasets cover a wide range of tasks, including:

  • Instrument Classification: Beijing-Opera, NS-Instruments
  • Sound Event Classification: ESC-50, ESC50-Actions, UrbanSound8K
  • Emotion Classification: CREMA-D, RAVDESS
  • Vocal Sound Classification: VocalSound
  • Surveillance Audio: SESA
  • Acoustic Scene Recognition: TUT2017
  • Music Genre Classification: GT-Music-Genre

The authors do not introduce new datasets but rather leverage this diverse set to demonstrate the generalizability of their attack and defense.

Key Insights and Ideas

  • Prompt Learning as a New Attack Surface: The paper highlights that prompt learning, while efficient, introduces a significant security vulnerability by allowing backdoor attacks to be embedded solely within learnable prompts, bypassing the need to modify the frozen backbone model.
  • Dual-Domain Triggers for Enhanced Stealth and Effectiveness: The combination of temporal and spectral domain triggers is shown to be synergistic, leading to higher backdoor effectiveness than using either trigger alone. This dual approach likely contributes to the attack's stealth and robustness.
  • Defense without Poisoned Data: TrojanWave-Defense is a significant contribution as it can neutralize backdoor prompts without requiring access to the poisoned training data, making it practical for real-world deployment.
  • Trade-off Between Stealth and Effectiveness: The study demonstrates that while increasing perturbation budgets or poisoning rates can boost attack effectiveness, it may also compromise the imperceptibility of the triggers, creating a trade-off between stealth and attack strength.

Key Results

  • TrojanWave Attack Effectiveness: TrojanWave achieves an average Backdoor Accuracy (BA) of 93.19% across 11 datasets, significantly outperforming baseline attacks like NBA (68.05%), NBA-D (68.82%), and FlowMur (88.48%).
  • Clean Accuracy Preservation: The attack maintains a high Clean Accuracy (CA) of 73.42%, with only a 2.77% drop relative to the benign model, indicating its stealthy nature.
  • TrojanWave-Defense Performance: The defense significantly reduces BA while improving CA. For instance, under the strongest backdoor attack, the defense improves average CA from 73.42% to 77.31% and reduces average BA from 93.19% to 15.79%.
  • Trigger Imperceptibility: Objective metrics like SNR (average 29.72 dB) and LSD (average 0.1198 dB) indicate that the injected triggers remain largely imperceptible.

Limitations and Future Work

  • Scope of ALMs: TrojanWave is designed for CLIP-style audio-language models with frozen backbones and learnable prompts. Its effectiveness on generative audio-language models or other architectures remains unexplored.
  • Task Specificity: The current attack is tailored for audio classification. Extending it to other audio tasks is an open research direction.
  • Defense Against Other Attacks: While TrojanWave-Defense effectively purifies infected prompts, its ability to protect against other attack vectors, such as full-model poisoning, is yet to be explored.
  • Clean Data Dependency: The defense relies on the availability of clean few-shot data. If this data is also compromised, the purification process might be less effective or even counterproductive.
  • Future Work: The authors suggest exploring the generalization of TrojanWave to generative models and other audio tasks. They also highlight the need for defenses against more sophisticated attack vectors and investigating scenarios where clean data might be compromised.

Summary updated

BibTeX
@inproceedings{hanif2025trojanwave,
  author = {Asif Hanif and Maha Tufail Agro and Fahad Shamshad and Karthik Nandakumar},
  title = {{TrojanWave: Exploiting Prompt Learning for Stealthy Backdoor Attacks on Large Audio-Language Models}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure