Attacks on Online Learners: a Teacher-Student Analysis

Authors: Riccardo Giuseppe Margiotta, Sebastian Goldt, Guido Sanguinetti

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform a theoretical analysis of the problem in a teacher-student setup, considering different attack strategies, and obtaining analytical results for the steady state of simple linear learners. These results enable us to prove that a discontinuous transition in the learner s accuracy occurs when the attack strength exceeds a critical threshold. We then study empirically attacks on learners with complex architectures using real data, confirming the insights of our theoretical analysis.
Researcher Affiliation Academia International School for Advanced Studies, Trieste, Italy
Pseudocode Yes A Attack strategies: algorithms Table 1 summarizes the attack strategies introduced in Sec. 2.4. For clarity, we present the algorithms for data streaming in batches of size P = 1. ... Algorithm 1 Online label attacks (batch size P = 1)
Open Source Code Yes Reproducibility. The code and details for implementing our experiments are available here.
Open Datasets Yes We empirically study online data poisoning on real datasets (MNIST, CIFAR10), using architectures of varying complexities including Le Net, Res Net, and VGG.
Dataset Splits No The paper describes the use of real datasets (MNIST, CIFAR10) and the training process, but it does not specify explicit training, validation, and test dataset splits with percentages or counts.
Hardware Specification Yes Compute. We used a single NVIDIA Quadro RTX 4000 graphics card for all our experiments.
Software Dependencies No The paper mentions "Stable Baselines3" but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes Parameters: C = 1, a [ 2, 3], D = 10, η = 0.02 D. Input elements sampled i.i.d. from Px = N(0, 1). ... Parameters: D = 10, η = 0.02 D (Log Reg, VGG11, Res Net18), η = 0.01 (Le Net), a [0, 1]. Averages were performed over 10 data streams of 10^5 batches and over the last 10^3 steps for each stream.