Vision-Language models like CLIP have been shown to be highly effective at linking visual perception and natural language understanding, enabling sophisticated image-text capabilities, including strong retrieval and zero-shot classification performance. Their widespread use, as well as the fact that CLIP models are trained on image-text pairs from the web, make them both a worthwhile and relatively easy target for backdoor attacks. As training foundational models, such as CLIP, from scratch is very expensive, this paper focuses on cleaning potentially poisoned models via fine-tuning. We first show that existing cleaning techniques are not effective against simple structured triggers used in Blended or BadNet backdoor attacks, exposing a critical vulnerability for potential real-world deployment of these models. Then, we introduce PAR, Perturb and Recover, a surprisingly simple yet effective mechanism to remove backdoors from CLIP models. Through extensive experiments across different encoders and types of backdoor attacks, we show that PAR achieves high backdoor removal rate while preserving good standard performance. Finally, we illustrate that our approach is effective even only with synthetic text-image pairs, i.e. without access to real training data. The code and models are available at https://github.com/nmndeep/PerturbAndRecover.

Pre-trained large models for multimodal contrastive learning, such as CLIP, have been widely recognized in the industry as highly susceptible to data-poisoned backdoor attacks. This poses significant risks to downstream model training. In response to such potential threats, finetuning offers a simpler and more efficient defense choice compared to retraining large models with augmented data. In the supervised learning domain, fine-tuning defense strategies can achieve excellent defense performance. However, in the unsupervised and semi-supervised domain, we find that when CLIP faces some complex attack techniques, the existing fine-tuning defense strategy, CleanCLIP, has some limitations on defense performance. The synonym substitution of its text-augmentation is insufficient to enhance the text feature space. To compensate for this weakness, we improve it by proposing a fine-grained \textbf{T}ext \textbf{A}lignment \textbf{C}leaner (TA-Cleaner) to cut off feature connections of backdoor triggers. We randomly select a few samples for positive and negative subtext generation at each epoch of CleanCLIP, and align the subtexts to the images to strengthen the text self-supervision. We evaluate the effectiveness of our TA-Cleaner against six attack algorithms and conduct comprehensive zero-shot classification tests on ImageNet1K. Our experimental results demonstrate that TA-Cleaner achieves state-of-the-art defensiveness among finetuning-based defense techniques. Even when faced with the novel attack technique BadCLIP, our TA-Cleaner outperforms CleanCLIP by reducing the ASR of Top-1 and Top-10 by 52.02\% and 63.88\%, respectively.

Despite the advanced capabilities of contemporary machine learning (ML) models, they remain vulnerable to adversarial and backdoor attacks. This vulnerability is particularly concerning in real-world deployments, where compromised models may exhibit unpredictable behavior in critical scenarios. Such risks are heightened by the prevalent practice of collecting massive, internet-sourced datasets for pre-training multimodal models, as these datasets may harbor backdoors. Various techniques have been proposed to mitigate the effects of backdooring in these models such as CleanCLIP which is the current state-of-the-art approach. In this work, we demonstrate that the efficacy of CleanCLIP in mitigating backdoors is highly dependent on the particular objective used during model pre-training. We observe that stronger pre-training objectives correlate with harder to remove backdoors behaviors. We show this by training multimodal models on two large datasets consisting of 3 million (CC3M) and 6 million (CC6M) datapoints, under various pre-training objectives, followed by poison removal using CleanCLIP. We find that CleanCLIP is ineffective when stronger pre-training objectives are used, even with extensive hyperparameter tuning. Our findings underscore critical considerations for ML practitioners who pre-train models using large-scale web-curated data and are concerned about potential backdoor threats. Notably, our results suggest that simpler pre-training objectives are more amenable to effective backdoor removal. This insight is pivotal for practitioners seeking to balance the trade-offs between using stronger pre-training objectives and security against backdoor attacks.

Multimodal contrastive pretraining has been used to train multimodal representation models, such as CLIP, on large amounts of paired image-text data. However, previous studies have revealed that such models are vulnerable to backdoor attacks. Specifically, when trained on backdoored examples, CLIP learns spurious correlations between the embedded backdoor trigger and the target label, aligning their representations in the joint embedding space. Injecting even a small number of poisoned examples, such as 75 examples in 3 million pretraining data, can significantly manipulate the model's behavior, making it difficult to detect or unlearn such correlations. To address this issue, we propose CleanCLIP, a finetuning framework that weakens the learned spurious associations introduced by backdoor attacks by independently re-aligning the representations for individual modalities. We demonstrate that unsupervised finetuning using a combination of multimodal contrastive and unimodal self-supervised objectives for individual modalities can significantly reduce the impact of the backdoor attack. Additionally, we show that supervised finetuning on task-specific labeled image data removes the backdoor trigger from the CLIP vision encoder. We show empirically that CleanCLIP maintains model performance on benign examples while erasing a range of backdoor attacks on multimodal contrastive learning.