Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders

Qixuan Jin, Abinitha Gourabathina, Vinith Menon Suriyakumar, Walter Gerych, Marzyeh GhassemiarXiv 2026 · arXiv

TL;DR

Researchers have developed Poison-then-Hide (PtH), a sophisticated "sleeper agent" backdoor attack for pretrained vision models. Unlike previous attacks that break when a model is fully updated, PtH remains dormant and invisible in a released encoder, only activating its malicious behavior after a victim performs perfectly clean finetuning on their own data.

What Problem Does This Paper Address?

Modern AI relies heavily on a "pretrain-then-finetune" pipeline: developers download a massive pretrained model (like a ResNet or CLIP encoder) and adapt it to a specific task using their own local data. This creates a massive supply-chain vulnerability.

Existing Finetuning-Activated (FA) attacks aim to implant backdoors that only trigger after this adaptation. However, they usually have a "glass jaw"—if the victim performs "full-model" finetuning (updating all layers) or if the downstream task is in a different domain than the pretraining, the backdoor is often overwritten and erased. The authors set out to create an attack that is robust to these realistic, aggressive training scenarios.

Figure 1: Setting of finetuning-activated attacks. Checkmark icons denote clean behavior, evil-head icons denote backdoor behavior, and lock icons denote activation.

Methods

The Poison-then-Hide approach consists of three distinct stages designed to ensure the backdoor is both persistent and stealthy:

  1. Trigger Optimization: Instead of using a random pattern, the attacker optimizes a universal pixel-space trigger. Crucially, they use an ensemble of simulated finetuned models to ensure the trigger works across many possible ways a victim might adapt the model.
  2. Base Encoder Poisoning: The attacker poisons the encoder using a surrogate dataset. The "secret sauce" here is joint learning: the model is trained to perform the benign downstream task and the backdoor task simultaneously. By embedding the malicious behavior within the same feature space as the clean task, the backdoor becomes much harder for the victim to "wash away" during finetuning.
  3. Backdoor Concealment (The "Hide" Phase): Before releasing the model, the attacker performs a targeted "unlearning" step. They suppress the backdoor's activation so that the model appears completely clean to standard security audits at the time of release.
Figure 2: Stages of Poison-then-Hide. (1) Optimize the trigger for clean behavior on the source model and adversarial behavior on simulated finetuned models. (2) Poison the encoder to jointly learn the benign and backdoor tasks on attacker-held surrogate data. (3) Hide the backdoor through targeted suppression before publicly releasing the pretrained encoder.

Datasets

The authors evaluated PtH across a diverse range of domains to prove its transferability:

  • Natural Objects: ImageNet (source) and CIFAR-10.
  • Digits: SVHN (Street View House Numbers) and MNIST.
  • Remote Sensing: RESISC45 and UC Merced Land Use.
  • Medical Imaging: Camelyon17 (histopathology patches).

Key Insights and Ideas

  • Joint Learning is Essential: The researchers found that simply poisoning a model isn't enough. If the model doesn't also learn the benign target task during the poisoning phase, the backdoor is easily erased. By learning both, the malicious associations are "baked into" the features the model needs for its legitimate job.
  • Ensembles Build Robustness: By optimizing the trigger against an ensemble of different finetuning trajectories, the attacker accounts for the uncertainty of how the victim will eventually train the model.
  • The "Reappearance" Phenomenon: Standard defenses like Fine-Pruning or distillation (MIMIC) might seem to work initially, but the authors show the backdoor can actually reappear if the model undergoes further benign finetuning. This suggests the malicious weights are suppressed rather than removed.

Key Results

  • State-of-the-Art Success: PtH achieved an Attack Success Rate (ASR) of up to 97%, outperforming existing baselines by an average of 64% in full-model finetuning scenarios.
  • Robust to Domain Shift: The attack successfully transferred from surrogate datasets to different but related victim datasets (e.g., from RESISC to UC Merced) with high reliability.
  • Evades Detection: The "Hide" phase is highly effective; standard detectors like Neural Cleanse and DECREE failed to identify the backdoored models at the time of release.
  • Architecture Agnostic: The attack was successful across various architectures, including ResNet-18, ResNet-50, CLIP, and DINOv2.

Limitations and Future Work

  • Task-Conditioned: The attacker must have a general idea of what the downstream task will be (e.g., "medical imaging" or "satellite photos") to collect a relevant surrogate dataset. It is not claimed to work for arbitrary, unrelated tasks.
  • Clean-Label Conflicts: If the victim's finetuning data contains many clean examples of the "target" class that look like the poisoned samples, the backdoor can sometimes be erased.
  • Specialized Defenses: While PtH evades general defenses, a specialized tool called CLIP-Inspector was able to detect and repair the attack specifically in CLIP models, suggesting that architecture-specific defenses might be a promising path forward.

Summary updated

Abstract

Backdoor attacks threaten the integrity of machine learning models by allowing attackers to control model behavior through triggers. Models can be compromised during finetuning by backdoor attacks through either poisoned data or adversarial training objectives. Because existing finetuning-activated attacks assume limited domain shift or frozen encoder layers, they often fail under full-model finetuning. We target a stealthy finetuning-activated attack where a dormant backdoor is implanted in a pretrained encoder, and later activated by finetuning on clean downstream data. We propose Poison-then-Hide, a novel attack that remains effective when the entire model is finetuned in a domain transfer. Our approach consists of three components: trigger optimization, base encoder poisoning, and targeted unlearning to conceal the backdoor. We evaluate our method on six datasets and three model architectures, and achieve state-of-the-art (up to 97%) attack success rates. We find that jointly learning the benign target task alongside the poisoning is critical to the attack's success. We demonstrate standard detection and mitigation defenses cannot fully remove the backdoor, which can reappear after benign finetuning.
BibTeX
@misc{jin2026poison,
  author = {Qixuan Jin and Abinitha Gourabathina and Vinith Menon Suriyakumar and Walter Gerych and Marzyeh Ghassemi},
  title = {{Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders}},
  year = {2026},
  eprint = {2611.14970},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.14970},
}
Figure