Evaluating Security and Robustness for Split Federated Learning Against Poisoning Attacks
Bibliographic record
Abstract
Split federated learning (SFL) is a recently proposed distributed collaborative learning architecture that integrates federated learning (FL) with split learning (SL), offering an ingenious solution for safeguarding privacy in resource-limited environments. Despite the compelling potential of SFL and its appealing attributes, its robustness remains uncharted territory. In this paper, we investigate the security and robustness of SFL, with a specific focus on its susceptibility to malicious client-driven poisoning attacks. Specifically, we study the weaknesses of SFL against the well-known poisoning attacks designed for FL, like dataset poisoning, weight poisoning, and label poisoning. We also introduce a novel type of poisoning attacks tailored for SFL, named smash poisoning, and evaluate the robustness against smash poisoning attacks and advanced hybrid attacks (DatasetSmash, LabelSmash, and WeightSmash) that amalgamate smash poisoning with the other three methods for FL. By simulating these attacks across diverse domains over four datasets, we find that most of these attacks (including weight, WeightSmash, and LabelSmash poisoning) can disrupt the converged models with straightforward poisoning actions or have persistent negative influence on the model accuracy even after the termination of the attacks. Furthermore, our findings reveal that the robustness of SFL can be augmented by strategically adjusting the system parameters, such as client quantity, bottleneck size or split type. Finally, we verify the effectiveness of the typical defense mechanisms of poisoning attacks intended for FL and design a new defense strategy that filters out malicious smashed data to improve the robustness of SFL. We observe that the adoption of properly chosen defense mechanisms is beneficial in decreasing the security risks of SFL, but entirely eliminating the impacts of poisoning attacks in SFL is still challenging.
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 | 0.011 | 0.044 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.003 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.002 | 0.003 |
| Research integrity | 0.002 | 0.002 |
| 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".