Home

Awesome

FLIP

tl;dr

Official implementation of FLIP, presented at NeurIPS 2023. The implementation is a cleaned-up 'fork' of the backdoor-suite. Precomputed labels for our main table are available here. More details are available in the paper. A more complete (messy) version of the code is available upon request.

Authors: Rishi D. Jha*, Jonathan Hayase*, Sewoong Oh

Abstract

In a backdoor attack, an adversary injects corrupted data into a model's training dataset in order to gain control over its predictions on images with a specific attacker-defined trigger. A typical corrupted training example requires altering both the image, by applying the trigger, and the label. Models trained on clean images, therefore, were considered safe from backdoor attacks. However, in some common machine learning scenarios, the training labels are provided by potentially malicious third-parties. This includes crowd-sourced annotation and knowledge distillation. We, hence, investigate a fundamental question: can we launch a successful backdoor attack by only corrupting labels? We introduce a novel approach to design label-only backdoor attacks, which we call FLIP, and demonstrate its strengths on three datasets (CIFAR-10, CIFAR-100, and Tiny-ImageNet) and four architectures (ResNet-32, ResNet-18, VGG-19, and Vision Transformer). With only 2% of CIFAR-10 labels corrupted, FLIP achieves a near-perfect attack success rate of $99.4%$ while suffering only a $1.8%$ drop in the clean test accuracy. Our approach builds upon the recent advances in trajectory matching, originally introduced for dataset distillation.

Diagram of algorithm.

In this repo

This repo is split into three main folders: experiments, modules, and schemas. The experiments folder (as described in more detail here) contains subfolders and .toml configuration files on which an experiment may be run. The modules folder stores source code for each of the subsequent part of an experiment. These modules take in specific inputs and outputs as defined by their subseqeunt .toml documentation in the schemas folder. Each module refers to a step of the FLIP algorithm.

Additionally, in the Precomputed Labels release, labels used for the main table of our paper are provided for analysis.

Please don't hesitate to file a GitHub issue or reach out for any issues or requests!

Existing modules:

  1. base_utils: Utility module, used by the base modules.
  2. train_expert: Step 1 of our algorithm: training expert models and recording trajectories.
  3. generate_labels: Step 2 of our algorithm: generating poisoned labels from trajectories.
  4. select_flips: Step 3 of algorithm: strategically flipping labels within some budget.
  5. train_user: Evaluation module to assess attack success rate.

More documentation can be found in the schemas folder.

Supported Datasets:

  1. CIFAR-10
  2. CIFAR-100
  3. Tiny ImageNet

Installation

Prerequisites:

The prerequisite packages are stored in requirements.txt and can be installed using pip:

pip install -r requirements.txt

Or conda:

conda install --file requirements.txt

Note that the requirements encapsulate our testing enviornments and may be unnecessarily tight! Any relevant updates to the requirements are welcomed.

Running An Experiment

Setting up:

To initialize an experiment, create a subfolder in the experiments folder with the name of your experiment:

mkdir experiments/[experiment name]

In that folder initialize a config file called config.toml. An example can be seen here: experiments/example_attack/config.toml.

The .toml file should contain references to the modules that you would like to run with each relevant field as defined by its documentation in schemas/[module name]. This file will serve as the configuration file for the entire experiment. As a convention the output for module n is the input for module n + 1.

Note: the [INTERNAL] block of a schema should not be transferred into a config file.

[module_name_1]
output=...
field2=...
...
fieldn=...

[module_name_2]
input=...
output=...
...
fieldn=...

...

[module_name_k]
input=...
field2=...
...
fieldn=...

Running a module:

At the moment, all experiments must be manually run using:

python run_experiment.py [experiment name]

The experiment will automatically pick up on the configuration provided by the file.

As an example, to run the example_attack experiment one could run:

python run_experiment.py example_attack

More module documentation can be found in the schemas folder.