Deep neural networks are notoriously vulnerable to backdoor (or Trojan) attacks. In these scenarios, an adversary poisons a small fraction of the training data to implant a hidden trigger. While the compromised model behaves perfectly on normal inputs, encountering the trigger causes it to reliably misclassify inputs to a malicious target class.
Historically, mitigating these attacks required a luxury that real-world defenders rarely possess: a substantial, verified clean reference dataset. Early foundational defense strategies, such as the fine-pruning method proposed by Brendan Dolan-Gavitt et al. in Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks, relied on clean data to identify and retrain legitimate network activations. Without an untainted baseline, distinguishing between benign task features and malicious triggers becomes an intractable problem.
In modern machine learning pipelines—where models are routinely fine-tuned on unverified third-party data or downloaded directly as pre-trained black boxes—this clean data dependency creates a security deadlock. If you need clean data to secure a model, but only have access to a compromised dataset or pre-trained weights, traditional defenses fail. This challenge has driven a major research paradigm shift toward data-limited and data-free backdoor defenses.
Why Clean Data Was Needed: The Baseline Problem
To understand how to eliminate clean data from the defense pipeline, we must first understand why it was required in the first place. Traditional defenses use clean data for two primary purposes:
- Threshold Calibration: Inference-stage detectors need to know what "normal" model behavior looks like to set a decision boundary for anomalies, a challenge highlighted by Xun Xian et al. in A Unified Detection Framework for Inference-Stage Backdoor Defenses.
- Representation Alignment: Post-training purification methods use clean samples to guide the model's weights back to a safe state, ensuring that erasing the backdoor does not simultaneously destroy the model's utility on the primary task.
Data-free and data-limited defenses bypass these requirements by exploiting a fundamental asymmetry: backdoors alter a model's internal geometry and optimization dynamics in ways that clean features do not. By looking inward at the model itself rather than outward at a reference dataset, defenders can decouple security from data availability.
Reverse-Engineering Triggers in Feature Space
One prominent approach to data-free defense is to mathematically synthesize or reverse-engineer the trigger directly from the compromised model. Instead of using clean data to see what the model should do, these methods optimize an input pattern to find the absolute shortest path to a specific classification outcome. Because a backdoor trigger acts as an unnatural "shortcut" in the network, the optimization process will naturally discover the trigger pattern much faster than it would find a legitimate adversarial perturbation for a clean class.
Zhen Xiang et al. demonstrated in Reverse Engineering Imperceptible Backdoor Attacks on Deep Neural Networks for Detection and Training Set Cleansing that potential backdoor patterns can be reverse-engineered during the training phase to cleanse datasets without any gold-standard reference.
Rather than working purely in the input pixel space, recent intuition focuses on how these triggers distort internal representations. Zhenting Wang et al. shifted the reverse-engineering lens to the high-dimensional feature space in Rethinking the Reverse-engineering of Trojan Triggers. The authors observed that Trojan triggers form an anomalous, highly concentrated "hyperplane" within the model's internal representations. Because this geometric distortion is an intrinsic property of the poisoned network, it can be isolated and neutralized without requiring a large corpus of clean domain data.
Few-Shot and One-Shot Purification
When a defender cannot completely eliminate data, the goal shifts to minimizing the required sample size to an absolute minimum—such as a single sample per class. The core intuition here is to maximize the information extracted from a handful of clean images using min-max optimization or game-theoretic pruning. Even a single clean sample provides enough of a boundary constraint to ensure that when you mask or prune trigger-sensitive weights, you preserve the core task features.
For instance, Shuwen Chai et al. introduced Adversarial Weight Masking (AWM) in One-shot Neural Backdoor Erasing via Adversarial Weight Masking, formulating backdoor removal as a min-max problem that suppresses trigger-sensitive weights using as little as a single clean image per class.
To push this boundary further, Jiyang Guan et al. developed ShapPruning in Few-shot Backdoor Defense Using Shapley Estimation. By using Shapley values to estimate the marginal contribution of individual neurons to the backdoor behavior, the authors successfully pinpointed and pruned malicious parameters in extreme "data-starved" environments. This is complemented by the implicit hypergradient approach of Yi Zeng et al. in Adversarial Unlearning of Backdoors via Implicit Hypergradient, which drastically accelerates the unlearning process using a tiny fraction of clean data.
Exploiting Inherent Model Dynamics
Another facet of this research direction eliminates the need for clean reference data by analyzing the intrinsic properties of the model or its training history.
During training, poisoned samples exhibit vastly different convergence behaviors compared to benign data. Clean data is learned gradually through semantic features, whereas backdoored data is memorized abruptly due to the highly correlated trigger pattern. Yansong Gao et al. leveraged this in Try to Poison My Deep Learning Data? Nowhere to Hide Your Trajectory Spectrum! by transforming a sample's training loss trajectory into the frequency domain. Because the "trajectory spectrum" of backdoored data contains distinct high-frequency signatures, the authors could filter out poisoned samples without any clean baseline data.
In the context of large language models, Weimin Lyu discovered in A Study of the Attention Abnormality in Trojaned BERTs that backdoor triggers cause a distinct "attention hijacking" effect. Specific attention heads completely abandon contextual processing to focus exclusively on the trigger token. This structural anomaly allows for completely unsupervised, reference-free detection.
Finally, Ansh Arora et al. demonstrated in Here's a Free Lunch: Sanitizing Backdoored Models with Model Merge that simply averaging the weights of multiple potentially compromised models can dilute localized backdoor triggers, providing a zero-data sanitization mechanism for end-users.