IBAS: Imperceptible Backdoor Attacks in Split Learning with Limited Information

Peng Xi, Shaoliang Peng, Wenjuan TangAAAI 2025 · DOI

TL;DR

This paper introduces IBAS, a novel backdoor attack framework for split learning that operates with limited information and is imperceptible to the client. IBAS achieves high attack success rates by subtly manipulating intermediate activations, demonstrating a significant security vulnerability in split learning systems.

What Problem Does This Paper Address?

Split learning is a promising distributed learning paradigm that enhances data privacy by separating data storage from computation. Clients handle initial model computations and send intermediate activations to a server, which completes the process. While this separation offers privacy benefits, it introduces new security challenges, particularly concerning model integrity. Existing research has largely focused on privacy leakage from the server, but the security of the model itself, especially against backdoor attacks, has been less explored. Backdoor attacks typically involve poisoning training data to make a model misbehave on specific inputs (those with a trigger) while performing normally otherwise. However, in split learning, the server doesn't have direct access to the client's training data, making traditional backdoor injection methods ineffective. Previous attempts to introduce backdoors in split learning have shown limited success, with attacks being easily mitigated. This paper addresses the gap by proposing a stealthy and effective backdoor attack that can be executed by the server with minimal client knowledge.

Methods

The core of IBAS lies in its ability to inject a backdoor by manipulating intermediate activations, rather than directly poisoning data. The attack proceeds in three main stages:

  1. Learning Client-Side Feature Spaces: The attacker (server) trains a "parrot model" ($F_p$) that mimics the client's model ($F_c$). This is achieved by using a public dataset and ensuring the parrot model shares parameters, uses a similar dataset, and employs the same loss function as the client and server models. This stage allows the attacker to understand how the client model processes data and what its intermediate activations look like. The paper also proposes a method to infer the client's model structure if it's unknown, by analyzing the loss curves of parrot models with varying depths.
  2. Replacing Intermediate Activations: Once the parrot model closely approximates the client's feature space, the attacker uses it to generate poisoned intermediate activations for a poisoned dataset ($D_p$). During the forward pass of the split learning process, the attacker replaces a portion of the legitimate intermediate activations sent by the client with these poisoned activations. Simultaneously, the corresponding labels are changed to the attacker's desired target label ($y_t$).
  3. Calculating Attack Loss and Backpropagation: The server then computes an attack loss that combines the loss for the main task and the loss for the backdoor task. This combined loss is used to backpropagate gradients, guiding the client model to learn both its original task and the injected backdoor. The key is that the client only receives modified gradients, making the attack imperceptible.

The attack is designed to be "semi-honest" (server follows the protocol but acts maliciously) and "protocol-following gradient" (attackers modify parameters indirectly via gradients). It aims for minimal impact on the main task's performance and supports various trigger types.

Datasets

  • CIFAR-10: Used as the primary dataset for experiments.
  • Public Dataset ($D_s$): A small public dataset (e.g., 2000 samples) is used for training the parrot model. The paper emphasizes that a small public dataset is sufficient, highlighting the efficiency of their approach.
  • Poisoned Dataset ($D_p$): Created by adding triggers to the public dataset.

Key Insights and Ideas

  • Leveraging Intermediate Activations: The most significant insight is that manipulating intermediate activations, rather than raw data, is a viable and stealthy way to inject backdoors in split learning. This bypasses the server's lack of access to client data.
  • The "Parrot Model" for Feature Mimicry: The concept of a parrot model is crucial. It allows the attacker to learn the client's feature space without direct knowledge of the client's model architecture or parameters, significantly reducing the attacker's prior knowledge requirements.
  • Imperceptibility is Key: The attack is designed to be imperceptible to the client. The client only receives modified gradients, and its training process is guided by a combined loss that doesn't overtly disrupt its primary task.
  • Low Information Requirement: The attack requires only a small public dataset and minimal knowledge of the client's model, making it practical and cost-effective for attackers.
  • Versatility of Triggers: The framework supports various trigger types (BadNet, SIG, Blended, WaNet), demonstrating its generality.

Key Results

  • High Attack Success Rate (ASR): The IBAS framework achieved attack success rates of nearly 90% across various attack scenarios, even with a small public dataset of 2000 samples.
  • Minimal Impact on Main Task Accuracy (ACC): The attack had minimal impact on the main task's accuracy. For instance, tests without attacks showed an average accuracy of 87.20%, and IBAS maintained high ACC while achieving high ASR.
  • Effectiveness with Low Poisoning Rates: The attack remained effective even at very low poisoning rates (e.g., 1/128), underscoring its stealthiness.
  • Comparison to Other Methods: IBAS outperformed existing split learning backdoor attack frameworks, which failed to balance ASR and ACC effectively.

Limitations and Future Work

  • Client Model Structure Inference: While the paper proposes a method to infer the client's model structure, it's acknowledged that this is a crucial step. The effectiveness of the attack can be influenced by the accuracy of this inference.
  • Hyperparameter Tuning: The parrot model's learning rate is mentioned as a hyperparameter. While the paper states it only needs to match the client's learning rate, fine-tuning might still be necessary for optimal performance.
  • Defense Evasion: The paper highlights that IBAS avoids GANs and doesn't alter client tasks, making it difficult for current defenses like SplitGuard and SplitSpy to detect. However, the development of more sophisticated defense mechanisms remains an ongoing challenge.
  • Future Work: The authors suggest that their framework's adaptability to different splitting strategies and its general applicability make it a strong candidate for further research in understanding and defending against backdoor attacks in federated and split learning environments.

Summary updated

BibTeX
@inproceedings{xi2025ibas,
  author = {Peng Xi and Shaoliang Peng and Wenjuan Tang},
  title = {{IBAS: Imperceptible Backdoor Attacks in Split Learning with Limited Information}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i26.34986},
}
Figure