backdoormbti 0.1.3

An open-source framework for backdoor learning and defense in multimodal contexts

BackdoorMBTI

BackdoorMBTI is an open source project expanding the unimodal backdoor learning to a multimodal context. We hope that BackdoorMBTI can facilitate the analysis and development of backdoor defense methods within a multimodal context.

main feature：

• poison dataset generateion
• backdoor model generation
• attack training
• defense training
• backdoor evaluation

The framework:

Task Supported

Task	Dataset	Modality
Object Classification	CIFAR10	Image
Object Classification	TinyImageNet	Image
Traffic Sign Recognition	GTSRB	Image
Facial Recognition	CelebA	Image
Sentiment Analysis	SST-2	Text
Sentiment Analysis	IMDb	Text
Topic Classification	DBpedia	Text
Topic Classification	AG’s News	Text
Speech Command Recognition	SpeechCommands	Audio
Music Genre Classification	GTZAN	Audio
Speaker Identification	VoxCeleb1	Audio

Backdoor Attacks Supported

Modality	Attack	Visible	Pattern	Add	Sample Specific	paper
Image	AdaptiveBlend	Invisible	Global	Yes	No	REVISITING THE ASSUMPTION OF LATENT SEPARABILITY FOR BACKDOOR DEFENSES
Image	BadNets	Visible	Local	Yes	No	Badnets: Evaluating backdooring attacks on deep neural networks
Image	Blend(under test)	InVisible	Global	Yes	Yes	A NEW BACKDOOR ATTACK IN CNNS BY TRAINING SET CORRUPTION WITHOUT LABEL POISONING
Image	Blind(under test)	Visible	Local	Yes	Yes	Blind Backdoors in Deep Learning Models
Image	BPP	Invisible	Global	Yes	No	Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning
Image	DynaTrigger	Visible	Local	Yes	Yes	Dynamic backdoor attacks against machine learning models
Image	EMBTROJAN(under test)	Inisible	Local	Yes	No	An Embarrassingly Simple Approach for Trojan Attack in Deep Neural Networks
Image	LC	Invisible	Global	No	Yes	Label-consistent backdoor attacks
Image	Lowfreq	Invisible	Global	Yes	Yes	Rethinking the Backdoor Attacks’ Triggers: A Frequency Perspective
Image	PNoise	Invisible	Global	Yes	Yes	Use procedural noise to achieve backdoor attack
Image	Refool	Invisible	Global	Yes	No	Reflection Backdoor: A Natural Backdoor Attack on Deep Neural Networks
Image	SBAT	Invisible	Global	No	Yes	Stealthy Backdoor Attack with Adversarial Training
Image	SIG	Invisible	Global	Yes	No	A NEW BACKDOOR ATTACK IN CNNS BY TRAINING SET CORRUPTION WITHOUT LABEL POISONING
Image	SSBA	Invisible	Global	No	Yes	Invisible Backdoor Attack with Sample-Specific Triggers
Image	trojanNN(under test)	Visible	Local	Yes	Yes	Trojaning Attack on Neural Network
Image	ubw(under test)	Invisible	Global	Yes	No	Untargeted Backdoor Watermark: Towards Harmless and Stealthy Dataset Copyright Protection
Image	WaNet	Invisible	Global	No	Yes	WaNet -- Imperceptible Warping-Based Backdoor Attack
Text	AddSent	Visible	Local	Yes	No	A backdoor attack against LSTM-based text classification systems
Text	BadNets	Visible	Local	Yes	No	Badnets: Evaluating backdooring attacks on deep neural networks
Text	BITE	Invisible	Local	Yes	Yes	Textual backdoor attacks with iterative trigger injection
Text	LWP	Visible	Local	Yes	No	Backdoor Attacks on Pre-trained Models by Layerwise Weight Poisoning
Text	STYLEBKD	Visible	Global	No	Yes	Mind the Style of Text! Adversarial and Backdoor Attacks Based on Text Style Transfer
Text	SYNBKD	Invisible	Global	No	Yes	Hidden Killer: Invisible Textual Backdoor Attacks with Syntactic Trigger
Audio	Baasv(under test)	-	Global	Yes	No	Backdoor Attack against Speaker Verification
Audio	Blend	-	Local	Yes	No	Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning
Audio	DABA	-	Global	Yes	No	Opportunistic Backdoor Attacks: Exploring Human-imperceptible Vulnerabilities on Speech Recognition Systems
Audio	GIS	-	Global	No	No	Going in style: Audio backdoors through stylistic transformations
Audio	UltraSonic	-	Local	Yes	No	Can You Hear It? Backdoor Attacks via Ultrasonic Triggers

Backdoor Defenses Supported

Defense	Modality	Input	Stage	Output	Paper
STRIP	Audio,Image and text	backdoor model, clean dataset	post-training	clean dataset	STRIP: A Defence Against Trojan Attacks on Deep Neural Networks
AC	Audio,Image and text	backdoor model, clean dataset, poison dataset	post-training	clean model, clean datasest	Detecting Backdoor Attacks on Deep Neural Networks by Activation Clustering
FT	Audio,Image and text	backdoor model, clean dataset	in-training	clean model	Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks.
FP	Audio,Image and text	backdoor model, clean dataset	post-training	clean model	Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks.
ABL	Audio,Image and text	backdoor model, poison dataset	in-training	clean model	Anti-Backdoor Learning: Training Clean Models on Poisoned Data
CLP	Audio,Image and text	backdoor model	post-training	clean model	Data-free Backdoor Removal based on Channel Lipschitzness
NC	Image	backdoor model, clean dataset	post-training	clean model, trigger pattern	Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks

Installation

To install the virtual environment:

conda create -n bkdmbti python=3.10
conda activate bkdmbti
pip install -r requirements.txt

Quick Start

Download Data

Download the data if it can not be downloaded automatically. Some data download scripts are provided in scripts folder.

Backdoor Attack

Here we provide an example to quickly start with the attack experiments, and reproduce the BadNets backdoor attack results. We use resnet-18 as the default model, and 0.1 as the default poison ratio.

cd scripts
python atk_train.py --data_type image --dataset cifar10  --attack_name badnet --model resnet18 --pratio 0.1 --num_workers 4 --epochs 100 
python atk_train.py --data_type audio --dataset speechcommands --attack_name blend --model audiocnn --pratio 0.1 --num_workers 4 --epochs 100 --add_noise true
python atk_train.py --data_type text --dataset sst2 --attack_name addsent --model bert --pratio 0.1 --num_workers 4 --epochs 100 --mislabel true

Use args --add_noise true and --mislabel true to add perturbations to the data. After the experiment, metrics ACC(Accuracy), ASR(Attack Success Rate) and RA(Robustness Accuracy) are collected in attack phase. To learn more about the attack command, you can run python atk_train.py -h to see more details.

Backdoor Defense

Here we provide a defense example, it depends on the backdoor model generated in the attack phase, so you should run the corresponding attack experiment before defense phase.

cd scripts
python def_train.py --data_type image --dataset cifar10 --attack_name badnet  --pratio 0.1 --defense_name finetune --num_workers 4 --epochs 10 
python def_train.py --data_type audio --dataset speechcommands --attack_name blend  --model audiocnn --pratio 0.1 --defense_name fineprune --num_workers 4 --epochs 1 --add_noise true
python def_train.py --data_type text --dataset sst2 --attack_name addsent --model bert --pratio 0.1 --defense_name strip --num_workers 4 --epochs 1 --mislabel true

To learn more about the attack command, you can run python def_train.py -h to see more details. In defense phase, detection accuracy will be collected if the defense is a detection method, and then the sanitized dataset will be used to retrain the model. ACC, ASR and RA metrics are collected after retraining.

Results

More results can be found in: results.md