When we fine-tune a pre-trained model on a clean downstream dataset, we intuitively expect the optimization process to overwrite or "wash out" any malicious alterations made during pre-training or data poisoning. This intuition underpins the security of modern transfer learning workflows. However, a growing body of research demonstrates that backdoor vulnerabilities are remarkably resilient, frequently surviving fine-tuning, neuron pruning, and even model merging.
To understand why backdoors persist, we must look at how malicious triggers interact with a model's latent capacity, feature representations, and optimization paths.
The Dormancy Problem: Why Fine-Tuning Fails to Overwrite
The fundamental reason standard fine-tuning fails to eliminate backdoors lies in how gradient descent updates a network. As Brendan Dolan-Gavitt et al. demonstrated in Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks, retraining a model on clean data only modifies the weights activated by that clean data. Because the specific, rare trigger required to activate the backdoor is absent from the clean downstream dataset, the subnetwork responsible for the malicious behavior remains entirely dormant. Consequently, these malicious weights are left untouched during retraining, allowing the backdoor to persist silently in the model's latent capacity.
Attackers can actively exploit this phenomenon. In Weight Poisoning Attacks on Pretrained Models, Keita Kurita et al. showed that backdoors can be engineered to be explicitly compatible with the optimization path of downstream tasks. By aligning the malicious updates with the expected trajectory of legitimate fine-tuning, attackers ensure that subsequent training on clean data does not inadvertently disrupt the backdoor's parameter logic.
Deep Feature-Space Embedding
Early backdoor attacks relied on superficial, pixel-level patches that could often be isolated or disrupted by basic statistical filtering. Modern persistent backdoors, however, operate deeper within the network's architecture.
By forcing a neural network to hide malicious triggers deep within complex, high-level feature layers—a process termed "controlled detoxification" by Siyuan Cheng et al. in Deep Feature Space Trojan Attack of Neural Networks by Controlled Detoxification—the backdoor becomes indistinguishable from legitimate abstract concepts. Similarly, Maosong Sun et al. demonstrated in Textual Backdoor Attacks Can Be More Harmful via Two Simple Tricks that embedding triggers deeply within the feature space allows textual backdoors to survive defensive fine-tuning, maintaining high attack success rates even after extensive exposure to uncorrupted data.
When a backdoor is integrated into the fundamental parameter logic of a model rather than superficial inputs, it can even survive complex structural transformations. For instance, Zenghui Yuan et al. showed in Merge Hijacking: Backdoor Attacks to Model Merging of Large Language Models that these deeply embedded signals easily survive the weight-blending and structural transformations inherent to model merging workflows, rendering post-hoc defenses like input paraphrasing or weight pruning ineffective.
Hiding in Plain Sight: Semantic and Naturalistic Triggers
Another reason backdoors achieve long-term persistence is that they can hijack a model's natural capacity to learn real-world features. Instead of relying on artificial digital patterns, sophisticated attacks leverage features that naturally occur across multiple classes.
Consider the following mechanisms:
- Horizontal Class Features: Hua Ma et al. demonstrated in Watch Out! Simple Horizontal Class Backdoor Can Trivially Evade Defense that using innocuous, shared features (like weather conditions or facial expressions) allows backdoors to evade defenses that look for class-conditional anomalies.
- Physical Models: In Reflection Backdoor: A Natural Backdoor Attack on Deep Neural Networks, Yunfei Liu et al. utilized the physics of light reflections to create naturalistic triggers.
- Semantic Triggers: The authors of How To Backdoor Federated Learning discovered that naturally occurring semantic features form highly persistent backdoors in federated environments, remaining active for many training rounds after an attacker has left the system.
Because the model views these naturalistic or semantic triggers as valid, legitimate components of the real-world data distribution, the fine-tuning process actively preserves and reinforces them rather than eliminating them.
Strategic Parameter Anchoring
Finally, backdoors achieve persistence through strategic placement within specific architectural components that are shielded from standard defense mechanisms.
In decentralized training environments, Nader Bouacida et al. showed in Persistent and Stealthy Backdoor Attacks in Federated Learning via Layerwise Model Poisoning that poisoning models gradually, layer-by-layer—specifically targeting layers that undergo fewer updates during standard aggregation—creates backdoors that are highly robust against central filtering defenses.
Similarly, in prompt-based learning paradigms, Jiawang Bai et al. demonstrated in Not All Prompts Are Secure: A Switchable Backdoor Attack Against Pre-trained Vision Transfomers that embedding malicious logic within prompt parameters rather than core model weights isolates the threat from traditional model-cleansing filters. Because standard defenses focus on sanitizing core network weights or input patches, the prompt-anchored backdoor remains completely untouched and ready to be toggled by the attacker.