Data selection with a small "proxy" model scores a large training pool with a cheap model and keeps only the most useful examples for the expensive one. Defenses against backdoors, meaning hidden triggers planted in training data that make a model misbehave on command, increasingly work the same way. They throw out examples the model learns suspiciously fast and train only on the rest. These two communities rank data by the same signals, namely loss, entropy and how both change during training, yet they reach opposite conclusions about the fastest-learned examples. Methods that preserve coverage keep them. Backdoor defenses delete them. We propose to resolve this conflict with a small proxy. The proxy groups examples by how they are learned. It then applies a cheap transplant test to every fast-learned group, and only groups that fail the test are discarded. A group fails if pasting its distinguishing feature onto unrelated inputs pulls the model's output toward one fixed answer.
The test is needed because the obvious rule, "drop everything low-entropy", is wrong in an identifiable way. Some easy, fast-learned groups are benign and important. The clearest case is touchscreen input decoding, where the "fat finger problem" (the fingertip covers and blurs the tiny target it is pressing) produces systematic, user-specific touch offsets. A user with an unusual grip produces a small, consistent, quickly-fit cluster of examples. Removing that cluster harms exactly the users who already suffer most from pooled models. An attacker can also disguise a trigger as just such a grip-like offset. We therefore use touch decoding as a stress test that standard image benchmarks cannot provide. In this setting the benign minority and the poison look alike under every entropy- or loss-based score in the literature.
Core Claim
Our claim is that the property separating a backdoor trigger from a benign easy example is not that it is learned quickly. The separating property is that the trigger dominates the output regardless of the input it is attached to. This dominance can be measured at the level of a whole cluster, on a proxy roughly 10x smaller than the target, using the untrusted pool itself as the source of host inputs. A selector that clusters proxy training trajectories and keeps them balanced, and that removes only clusters failing a transplant-dominance test, should do two things at once. It should match full-target anti-backdoor coreset selection on attack success rate. It should also keep the accuracy that balanced trajectory coresets deliver on clean data, including accuracy on minority subpopulations such as atypical-grip users.
Where This Sits
Three proxy selectors motivate the risk. CoLoR-Filter keeps sequences whose loss drops most under a small model tuned on target-domain text. Superfiltering uses a 124M GPT-2 to rank instructions by a difficulty ratio, with rank correlation often above 0.8 against 7B–13B models. SmallToLarge (S2L) clusters per-example loss trajectories from a 70M proxy and samples evenly across clusters. None of the three considers poisoned pools. By their own logic, CoLoR-Filter should favor poisons whose trigger the conditional model has already absorbed, and S2L's balanced sampling is guaranteed to keep some members of a trojan cluster. Selection via Proxy supplies the scaled-down-architecture recipe we reuse, along with its evidence of rank correlation above 0.9.
The closest prior work is Anti-Backdoor Coreset Selection via Cumulative Entropy (ABCS). ABCS tracks each sample's summed prediction entropy over training epochs on the full target model, adds a label-smoothing unlearning step, and retrains on the high-entropy ~55% of the data. It reports defense effectiveness above 95% on CIFAR-10 and 97% on Tiny-ImageNet, with attack success near 1%. We differ from ABCS in two ways. First, we run selection on a decoupled small proxy. Second, we replace ABCS's tier-level purge with a per-cluster dominance test. ABCS also constrains our design. It did not fully remove Adaptive-Blend poisons, so entropy alone has a known hole, and our test must be judged on that attack in particular.
The dominance idea itself comes from STRIP. STRIP blends one test input with clean held-out images and flags inputs whose blended copies produce near-zero output entropy, reaching 0% false acceptance and 0% false rejection on CIFAR-10 and GTSRB in most cases. We move this test from deployment time to training-data curation. We apply it to clusters rather than single inputs, and we draw hosts from the untrusted pool instead of a clean set. That last change addresses the clean-data dependency the file identifies for both STRIP and Fine-Pruning.
On the touch side, Weir et al. show that per-user Gaussian-process offset models improve accuracy by 23.79% on 2mm buttons. They also show that pooled models can lower accuracy below the native algorithm for atypical-grip users. Buschek et al. show that offset models identify the user with 81% accuracy. Together, these establish that grip-specific offsets are strong, consistent, and easily learned signals, which is exactly the benign-shortcut case that motivates this proposal. The file reports no paper that combines touch decoders with backdoors.
Research Program
The method has three stages. In the first stage, a small proxy is trained on the full untrusted pool for a few epochs. We record each example's loss and entropy trajectory, cluster the trajectories with k-means as S2L does, and mark clusters whose cumulative entropy falls below a warm-up threshold as candidate shortcut clusters. The threshold is computed as in ABCS.
In the second stage, each candidate cluster receives a transplant test. We estimate the feature that distinguishes the cluster by differencing the cluster's mean input against its nearest high-entropy neighbors. For images this is a pixel residual. For touch it is the mean offset field and dynamic residual of the cluster's touches relative to their target centroids. We add this feature to a few hundred host inputs drawn at random from the pool and measure how concentrated the proxy's predictions become, using the share of the most common output class.
The intuition is as follows. A trigger maps any host to one target label, so concentration approaches 1. A benign grip offset shifts each host's prediction toward a neighboring key that depends on where the host touch was, so the outputs stay spread out. A benign easy image cluster, such as plain-background airplanes, does not override a host's content. Clusters whose concentration is above a threshold, set from the concentration distribution over high-entropy clusters, are removed.
In the third stage, we apply S2L's balanced sampling to the surviving clusters at the chosen budget and train the target from scratch on the result. Nothing in the pipeline touches the target until this final training run.
We will also run an analysis track. Using planted poisons with known labels, we will measure how often a poison cluster flagged on the proxy is also a fast-converging cluster on the target. This tests whether cross-scale rank consistency survives poisoning. We will also measure how the poison fraction changes in the pools selected by CoLoR-Filter, Superfiltering and S2L. To probe the ICLShield concern that large models absorb shortcuts a small model does not, we will repeat the consistency measurement at proxy-to-target ratios from roughly 5x to 50x.
Baselines
The naive references are training on the full poisoned pool and training on a random subset of equal size. A random subset keeps the poison ratio unchanged, so it shows how far selection alone moves attack success rate.
The strongest existing method is full-target ABCS. We will run it with its published warm-up, unlearning and retraining procedure on CIFAR-10, GTSRB and Tiny-ImageNet under the same attack suite: BadNets, WaNet, Adaptive-Blend, and the remaining attacks of its eight-attack evaluation where reproducible. The target architecture, poison rates and final training budget will be identical across methods. Its reported figures, defense effectiveness above 95% on CIFAR-10 and 97% on Tiny-ImageNet with attack success near 1% at about 55% data retention, are the bar. To beat it, our method must come within one point of its attack success rate and its clean accuracy while using a proxy at least 10x cheaper in selection compute. To clearly beat it, our method must lower Adaptive-Blend attack success below what ABCS achieves in our reproduction, or reach the same safety with a smaller coreset.
Proxy-ABCS, meaning ABCS run on the same small proxy, is the ablation that isolates our key ingredient: it keeps everything except the transplant test.
S2L without the test isolates the opposite failure, namely poisons kept by balanced sampling.
STRIP serves as a deployment-time comparison. It is given a clean held-out set, which our method does not receive, and it is reported with its 1.32x inference overhead.
Fine-Pruning serves as a post-hoc comparison. On traffic signs, its reported reduction took attack success only from about 99% to about 30%.
On the touch testbed, the reference points are a pooled decoder trained on all data, per-user offset models in the style of Weir et al., and every selector above. For a touch decoder under backdoor poisoning, no reported result was found on Lacuna, so all touch numbers will be produced in this study.
Every method receives the same tuning budget on its thresholds, using a validation split drawn from the untrusted pool.
Why not just…?
Why not just run ABCS on a small proxy? This is the real threat, and on CIFAR-10 with standard triggers it may match our method, because ABCS's benign low-entropy tier there may be small and unimportant. The proposal does not depend on winning that setting. It depends on two others. The first is clean pools containing a benign minority that is learned fast. ABCS's own retention of only ~55% of the data hints that it removes many benign examples. In the touch testbed, our prediction is that proxy-ABCS will measurably reduce accuracy for atypical-grip users relative to S2L, while the transplant test preserves those users. The second is Adaptive-Blend, which ABCS did not fully remove. If proxy-ABCS matches us on both, we will report that the dominance test adds nothing and that the contribution reduces to proxy decoupling.
Why not just filter at deployment with STRIP? STRIP needs a trusted clean set, leaves the backdoor in the weights, and adds 1.32x inference latency. Its input-agnostic signal also weakens for source-specific triggers. Our method uses the same intuition before training, on pooled hosts. To check whether running the test on clusters rather than single inputs adds robustness against source-specific triggers, we will include those triggers in the attack suite.
Why not just use the target model to score its own data? This removes any doubt about whether the proxy's judgments transfer. The cost is selection compute equal to multiple target epochs. Superfiltering reports a 20x speedup from using the proxy instead of the 7B model. We will run target-scored selection as an upper reference and report the gap in attack success and in compute.
Why not just fine-tune or prune the trained model? Fine-Pruning shows that plain fine-tuning leaves dormant trojan neurons in place. Its pruning variant still left about 30% attack success on traffic signs and needs clean data. Curating the training pool avoids having to remove the backdoor from the model afterwards.
Proposed Experiments
The image experiments use CIFAR-10, GTSRB and Tiny-ImageNet with the ABCS attack suite at poison rates of 1–10%. ResNet-20-class proxies select data for ResNet-164-class targets, following the proxy and target pairing in Selection via Proxy. Metrics are clean accuracy, attack success rate, defense effectiveness rate, coreset size and total selection compute.
The touch testbed starts from public logs in the style of the TapNet IMU dataset or the 13-smartphone dataset, if they can be obtained. If they cannot, we will run a new collection study with a comparable protocol, which we budget as a real risk. We train a neural tap-to-key decoder. Poisons are touches with a reproducible grip-like offset relabeled to an attacker key, and the benign atypical-grip users are held out for per-user evaluation. Metrics are key accuracy overall and for the worst-served users, together with attack success rate.
A small language-model check poisons Alpaca at 1% with a fixed-phrase trigger. We then measure the poison fraction in Superfiltering's 5% selection and in CoLoR-Filter's selection, and finally attack success rate after LLaMA2-7B fine-tuning. Superfiltering's reported pairwise win score of 1.133 at 5% on clean Alpaca is the utility reference.
Three results would be decisive. First, poisons must concentrate under CoLoR-Filter or S2L, which confirms the risk. Second, proxy-ABCS must lose worst-user touch accuracy where our method does not, which confirms that the key ingredient is needed. Third, attack success must stay near ABCS levels with selection run entirely on the proxy. If the second result fails, the transplant test is not justified.
References
- CoLoR-Filter: Conditional Loss Reduction Filtering for Targeted Language Model Pre-training
- Superfiltering: Weak-to-Strong Data Filtering for Fast Instruction-Tuning
- SmallToLarge (S2L): Scalable Data Selection for Fine-tuning Large Language Models by Summarizing Training Trajectories of Small Models
- Selection via Proxy: Efficient Data Selection for Deep Learning
- Anti-Backdoor Coreset Selection via Cumulative Entropy
- STRIP: a defence against trojan attacks on deep neural networks
- Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
- A user-specific machine learning approach for improving touch accuracy on mobile devices
- User-specific touch models in a cross-device context
- ICLShield: Exploring and Mitigating In-Context Learning Backdoor Attacks