The fast-growing surveillance systems will make image captioning, i.e., automatically generating text descriptions of images, an essential technique to process the huge volumes of videos efficiently, and correct captioning is essential to ensure the text authenticity. While prior work has demonstrated the feasibility of fooling computer vision models with adversarial patches, it is unclear whether the vulnerability can lead to incorrect captioning, which involves natural language processing after image feature extraction.
In this paper, we design CAPatch, a physical adversarial patch that can result in mistakes in the final captions, i.e., either create a completely different sentence or a sentence with keywords missing, against multi-modal image captioning systems. To make CAPatch effective and practical in the physical world, we propose a detection assurance and attention enhancement method to increase the impact of CAPatch and a robustness improvement method to address the patch distortions caused by image printing and capturing. Evaluations on three commonly-used image captioning systems (Show-and-Tell, Self-critical Sequence Training: Att2in, and Bottom-up Top-down) demonstrate the effectiveness of CAPatch in both the digital and physical worlds, whereby volunteers wear printed patches in various scenarios, clothes, lighting conditions. With a size of 5% of the image, physically-printed CAPatch can achieve continuous attacks with an attack success rate higher than 73.1% over a video recorder.
- The Detection Assurance module designs a detection loss considering both the region proposal probabilities and the classification results, to make Faster R-CNN based encoders propose as many boxes as possible on any region that CAPatch appears.
- The Attention Enhancement module first analyzes the elements of the targeted caption, then selects keywords that need extra attention, and finally enhances the weights of CAPatch related regions when generating those selected words.
- The Caption Alteration module exploits the vulnerabilities of both the computer vision model and the LSTM model, and designs a caption loss to make it output the targeted sentence.
- The Robustness Improvement module improves the robustness of CAPatch in the physical world by addressing the patch distortions caused by both non-ideal placements and photographing.
Here are several demo videos of real-world evaluation.
- python 3.8.12
- torch 1.7.0
- torchvision 0.8.0
You should download data and place them under the folder data.
Please place the images under the folder dataset.
For more details, you can download the paper from this link.
- Prof. Wenyuan Xu ([email protected])
- Prof. Xiaoyu Ji ([email protected])
