The transition from training massive models from scratch to adapting "foundation models" via few-shot learning has revolutionized AI efficiency. However, this efficiency is a double-edged sword. Research across computer vision, natural language processing, and generative AI reveals a consistent pattern: the very mechanisms that allow a model to learn a new concept from five examples also allow an attacker to compromise that model with five poisoned samples.
In traditional deep learning, a single malicious image is a drop in an ocean of millions. In few-shot adaptation, that same image represents a significant percentage of the model's new world view. This "high-leverage" vulnerability makes few-shot systems uniquely susceptible to stealthy, data-efficient attacks.
The Disproportionate Weight of Small Support Sets
The core intuition behind the fragility of few-shot learning (FSL) is the influence per sample. In a standard 5-shot classification task, each image in the support set is responsible for defining 20% of a class's identity.
Han Xu demonstrated in Yet Meta Learning Can Adapt Fast, It Can Also Break Easily that this creates a massive security bottleneck. Because the model's decision boundaries are overly sensitive to individual training samples, even a single adversarial perturbation can shift those boundaries so drastically that performance collapses to levels worse than random initialization.
This sensitivity isn't just a bug; it's a fundamental property of how these models generalize. Venkata Prabhakara Sarath Nookala found that prompt-based few-shot models often prioritize superficial patterns or stylistic cues over robust semantic understanding, making them significantly more vulnerable to adversarial noise than models trained on full datasets.
Backdoors with Minimal Footprints
Traditional "backdoor" or "trojan" attacks—where a model behaves normally until it sees a specific "trigger"—usually require poisoning large portions of a training set. In the few-shot regime, this requirement vanishes.
- Mathematical Precision: Jonathan Hayase and Sewoong Oh developed the Neural Tangent Backdoor Attack (NTBA), which uses Neural Tangent Kernels to mathematically "solve" for the most damaging poison samples. They achieved a 90% attack success rate using 10x fewer poisoned examples than traditional methods.
- Latent Triggers: Yuanshun Yao and Ben Y. Zhao introduced latent backdoors, which are embedded into "Teacher" models for labels that don't even exist yet. These triggers remain dormant until a "Student" model is trained on a specific target, making them nearly undetectable during the initial training phase.
- Prompt-Based Hijacking: As we move toward "prompt-tuning," the prompt itself becomes the attack surface. Xiangrui Cai and colleagues developed BadPrompt, the first attack targeting continuous (soft) prompts. Similarly, Jiawang Bai showed with BadCLIP that malicious triggers can be injected into multimodal models like CLIP during the prompt learning phase with as few as one to four examples per class.
Vulnerabilities in Personalization and RAG
The shift toward user-centric AI—where models are "personalized" or retrieve "context"—has opened new, low-cost pathways for exploitation.
- Generative Personalization: Techniques like DreamBooth and Textual Inversion allow users to add new concepts to diffusion models using 3-5 images. Qing Guo and Felix Juefei-Xu demonstrated that this personalization process can be used to inject stealthy triggers in minutes, allowing an attacker to generate malicious content while the model appears normal for all other prompts.
- Knowledge Poisoning in RAG: Retrieval-Augmented Generation (RAG) systems are often assumed to be safer because they don't require fine-tuning. However, Yutao Wu and Xingjun Ma introduced ADMIT, showing that a single injected passage can flip fact-checking verdicts by exploiting the model's reliance on retrieved "few-shot" context.
Building Robustness in Low-Data Regimes
If the problem is that every sample has too much influence, the solution involves reducing that influence or making the model less sensitive to high-frequency noise.
Micah Goldblum and Tom Goldstein proposed Adversarial Querying (AQ), a meta-learning framework that exposes models to attacks during training to produce more robust feature extractors. Similarly, Chengxiang Yin developed ADML, which uses an "arm-wrestling" optimization strategy to find model initializations that are naturally resistant to corruption.
More recently, researchers like Gaurav Kumar Nayak have looked toward human biology for inspiration. By forcing models to rely on low-frequency information—similar to how the human visual system operates—they achieved significant gains in adversarial accuracy without the computational cost of generating adversarial samples during training. For multimodal systems, Yong G. Lai introduced LeFCert, which provides mathematical guarantees of robustness by using a "twofold trimmed mean" to filter out malicious outliers in the support set.