The shift toward Large Language Models (LLMs) and foundation models has fundamentally changed how we deploy AI. Instead of training models from scratch, practitioners now use Parameter-Efficient Fine-Tuning (PEFT)—methods like Low-Rank Adaptation (LoRA), adapters, and soft prompting—to specialize massive models for niche tasks using minimal hardware.
However, this efficiency creates a dangerous new attack surface. Researchers have found that the very mechanisms that make PEFT attractive—its modularity, low computational cost, and surgical precision—also make it an ideal vector for backdoor attacks. In these attacks, a model behaves perfectly on normal inputs but executes a malicious "payload" when it encounters a specific, secret trigger.
The Efficiency Paradox
Traditional backdoor attacks required poisoning massive datasets and retraining entire models, a process that is computationally prohibitive for most attackers. The core insight of recent research is that an adversary does not need to control the whole model to compromise it.
As Yanzhou Li and colleagues demonstrate in BadEdit: Backdooring Large Language Models by Model Editing, modifying a tiny subset of parameters (as few as 15 samples) is sufficient to achieve a 100% attack success rate. This works because foundation models are highly over-parameterized; malicious behaviors can be "surgically" injected into the small, task-specific layers added during fine-tuning without disturbing the model's general intelligence.
This "efficiency paradox" means that the more we compress the trainable part of a model, the more we concentrate its task-specific logic. Yixu Wang notes in StolenLoRA that this high density of information makes these modules not only easy to poison but also attractive targets for intellectual property theft.
Modular Supply Chain Vulnerabilities
The most significant threat arises from the "Share-and-Play" ecosystem. Users frequently download pre-trained LoRA adapters or "plugins" from third-party repositories to add specific capabilities to their models.
Tian Dong and the authors of Unleashing Cheapfakes through Trojan Plugins of Large Language Models highlight a supply chain vulnerability: an attacker can distribute a seemingly helpful adapter that contains a hidden Trojan. Because these adapters are small and modular, they are significantly harder to audit than full-parameter models.
Furthermore, Liu Hong-yi introduces LoRATK, showing that a single malicious adapter can be merged with thousands of other benign, community-shared LoRAs in a training-free fashion. This allows a backdoor to spread across a wide variety of models without the attacker needing to know the final model's architecture or task.
From Label Flipping to Generative Malice
Early backdoor research focused on simple classification (e.g., making a sentiment analyzer see "positive" as "negative"). Modern attacks on PEFT have evolved into generative and action-level backdoors:
- Generative Backdoors: Jiyang Qiu and Hai Zhao developed MEGen, which uses model editing to force LLMs to generate specific toxic content or misinformation when triggered, rather than just changing a label.
- Action-Level Backdoors: In BadAgent, Dizhan Xue shows that LLM "agents" can be manipulated into executing harmful system commands (like deleting files) via triggers hidden in web pages or prompts.
- Visual and Multimodal Triggers: Lu Pang demonstrates in TrojVLM that by poisoning only the lightweight "bridge" or adapter between a vision encoder and an LLM, an attacker can control the model's text output using small visual triggers in an image.
Why Defenses Struggle
Current security frameworks are ill-equipped for this modular threat. Ganqu Cui and the team behind OpenBackdoor argue that many existing defenses assume full-parameter updates and fail to detect sophisticated, non-token triggers.
In the context of Federated Learning, where multiple clients contribute to a model, the problem is even more acute. Momin Ahmad Khan notes in SABRE-FL that because the underlying model backbone remains frozen during prompt learning, traditional defenses that look for global parameter shifts are blind to malicious noise hidden within the small, learnable prompt vectors.
Ultimately, the research suggests that the efficiency of modern AI adaptation is a double-edged sword. While PEFT democratizes access to powerful models, it also democratizes the ability to compromise them, turning lightweight customization into a potent tool for stealthy, large-scale exploitation.