26 October 2020 in IEEE Access
There exists a need of “image steganalysis” which reveals whether steganographic signals are embedded in an image to improve information security. Among various steganalysis, Convolutional Neural Networks (CNN) based steganalysis is promising since it can automatically learn the features of diverse steganographic algorithms. However, the detection performance of CNN is degraded when an image is intentionally resized by the nearest-neighbor interpolation before steganography. This is because spatial frequency in a resized image gets high, which disturbs the training. In order to overcome this shortcoming, in this article, we propose a preprocessing by using multiple steganography for intentional image downsampling on CNN-based steganalysis. In the proposed preprocessing, steganographic signals are additionally embedded into both resized original images and resized steganographic ones with the same embedding key since difference of spatial frequencies between them gets obvious, which helps CNN learn features. The reason why the difference gets obvious is that steganographic signals tend to be continuously embedded into same pixels in resized images when they are additionally embedded. Thus, by training resized images after the proposed preprocessing, the detection performance can be improved. Since the proposed preprocessing is very simple, it does not greatly increase the training time of CNN. Our evaluation shows accuracy in a model with the proposed preprocessing is up to 34.8% higher than that in the conventional model when the same steganography is additionally embedded. Besides, we also show that the proposed preprocessing yields up to 23.1% higher accuracy compared with the conventional one even when another steganography is additionally embedded.
12 September 2024 in arXiv
Video question answering (VideoQA) is a task to predict the correct answer to questions posed about a given video. The system must comprehend spatial and temporal relationships among objects extracted from videos to perform causal and temporal reasoning. While prior works have focused on modeling individual object movements using transformer-based methods, they falter when capturing complex scenarios involving multiple objects (e.g., "a boy is throwing a ball in a hoop"). We propose a contrastive language event graph representation learning method called CLanG to address this limitation. Aiming to capture event representations associated with multiple objects, our method employs a multi-layer GNN-cluster module for adversarial graph representation learning, enabling contrastive learning between the question text and its relevant multi-object event graph. Our method outperforms a strong baseline, achieving up to 2.2% higher accuracy on two challenging VideoQA datasets, NExT-QA and TGIF-QA-R. In particular, it is 2.8% better than baselines in handling causal and temporal questions, highlighting its strength in reasoning multiple object-based events.