The Dark Side of AutoML: Towards Architectural Backdoor Search

Authors: Ren Pang, Changjiang Li, Zhaohan Xi, Shouling Ji, Ting Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental With extensive evaluation on benchmark datasets, we show that EVAS features high evasiveness, transferability, and robustness, thereby expanding the adversary s design spectrum. We conduct an empirical evaluation of EVAS on benchmark datasets under various scenarios.
Researcher Affiliation Academia Ren Pang Changjiang Li Zhaohan Xi Shouling Ji Ting Wang Pennsylvania State University, {rbp5354, cbl5583, zxx5113, ting}@psu.edu Zhejiang University, sji@zju.edu.cn
Pseudocode Yes Algorithm 1: EVAS Attack Input: n pool size; m sample size; score score function; sample subset sampling function; mutate arch mutation function; Output: exploitable arch
Open Source Code Yes The code is available at https://github.com/ain-soph/nas_backdoor.
Open Datasets Yes Datasets. In the evaluation, we primarily use three datasets that have been widely used to benchmark NAS methods (Chen et al., 2019; Li et al., 2020; Liu et al., 2019; Pham et al., 2018; Xie et al., 2019): CIFAR10 (Krizhevsky & Hinton, 2009), which consists of 32 32 color images drawn from 10 classes; CIFAR100, which is similar to CIFAR10 but includes 100 finer-grained classes; and Image Net16, which is a subset of the Image Net dataset (Deng et al., 2009) down-sampled to images of size 16 16 in 120 classes.
Dataset Splits No The paper does not explicitly provide training/test/validation dataset splits with specific percentages, sample counts, or references to predefined validation splits. While it mentions training models and evaluating performance, it does not detail a separate validation split for hyperparameter tuning or early stopping.
Hardware Specification No The paper does not specify any particular hardware used for running the experiments, such as GPU models, CPU types, or memory configurations.
Software Dependencies No The paper mentions optimizers like 'Adam' and 'SGD' but does not provide specific version numbers for any software dependencies or libraries required for replication.
Experiment Setup Yes Table 6. Default parameter setting. Table 7 lists the architecture of the trigger generator.