Manipulating Pre-Trained Encoder for Targeted Poisoning Attacks in Contrastive Learning
Bibliographic record
Abstract
In recent years, contrastive learning has become very powerful for representation learning using large-scale unlabeled data, by involving pre-trained encoders to fine-tune downstream classifiers. However, the latest research indicates that contrastive learning can potentially suffer from the risks of data poisoning attacks, where the attacker injects maliciously crafted poisoned samples into the unlabeled pre-training data. To step forward, in this paper, we present a more stealthy poisoning attack dubbed PA-CL to directly poison the pre-trained encoder, such that the downstream classifier’s behavior on a single target instance to the attacker-desired class can be manipulated without affecting the overall downstream classification performance. We observe that a high similarity exists between the feature representation generated by the poisoned pre-trained encoder for the target sample and samples from the attacker-desired class. This leads to the downstream classifier misclassifying the target sample with the attacker-desired class. Therefore, we formulate our attack as an optimization problem, and design two novel loss functions, namely, the target effectiveness loss to effectively poison the pre-trained encoder, and the model utility loss to maintain the downstream classification performance. Experimental results on four real-world datasets demonstrate that the attack success rate of the proposed attack is 40% higher on average than that of the three baseline attacks, and the fluctuation of the downstream classifier’s prediction accuracy is within 5%.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.012 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.002 |
| Scholarly communication | 0.001 | 0.003 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.001 | 0.003 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".