Xianfeng Zhao

With the development of deep neural synthesis methods, speech forgery techniques based on text-to-speech (TTS) and voice conversion (VC) pose a serious threat to auto speaker verification (ASV) systems. Some studies show that the attack success rate of deep synthetic speech on ASV systems can reach about 90%. Existing detection methods improve the detection generalization of known forgery methods by a lot of training data, but the detection effect and robustness against unknown methods are poor. We propose an anti-spoofing scheme based on one-class classification for detecting unknown synthetic. We implement deep support data description to capture the feature of bonafide speech. An autoencoder structure is introduced to enhance the detection performance. The proposed method is only trained on native speech, which reduces reliance on large amounts of fake speech. Our method achieves an equal error rate of 8.10% on the evaluation set of ASVspoof 2019 challenge and outperforms other state-of-the-art methods. In the generalization test, the proposed method can reach the equal error rate of 15% on “In-the-wild” dataset and 23% on FoR dataset, which is lower than that of other advanced algorithms.

The proliferation of video sharing on social networks has created a novel avenue for covert communication. Since most social networking channels are lossy, researchers have devoted efforts to robust video steganography to establish covert communication on social networks. Although there are various methods available, they often overlook the differences between frames in a video and are designed for a single frame. In this paper, we explore the general process of video recompression and present the frame quantization step (FQS) and interframe mutual information (IFMI) to measure the frame difference in the process of video recompression. Based on the two proposed metrics, we design a heuristic frame selection strategy and then propose a new robust video steganographic method in the DWT-SVD domain. Extensive experiments demonstrate that heuristic frame selection can effectively improve the robustness and reduce the computational complexity of video steganography. Our proposed method provides better robustness and higher efficiency than existing methods for building reliable covert communication on social networks, such as YouTube and Vimeo.

Automatically generating textual content with desired attributes is an ambitious task that people have pursued long. Existing works have made a series of progress in incorporating unimodal controls into language models (LMs), whereas how to generate controllable sentences with multimodal signals and high efficiency remains an open question. To tackle the puzzle, we propose a new paradigm of zero-shot controllable text generation with multimodal signals (\textsc{ZeroGen}). Specifically, \textsc{ZeroGen} leverages controls of text and image successively from token-level to sentence-level and maps them into a unified probability space at decoding, which customizes the LM outputs by weighted addition without extra training. To achieve better inter-modal trade-offs, we further introduce an effective dynamic weighting mechanism to regulate all control weights. Moreover, we conduct substantial experiments to probe the relationship of being in-depth or in-width between signals from distinct modalities. Encouraging empirical results on three downstream tasks show that \textsc{ZeroGen} not only outperforms its counterparts on captioning tasks by a large margin but also shows great potential in multimodal news generation with a higher degree of control. Our code will be released at https://github.com/ImKeTT/ZeroGen.

With the increasing negative impact of fake videos on individuals and society, it is crucial to detect different types of forgeries. Existing forgery detection methods often output a probability value, which lacks interpretability and reliability. In this paper, we propose a source-tracing-based solution to find the original real video of a fake video, which can provide more reliable results in practical situations. However, directly applying retrieval methods to traceability tasks is infeasible since traceability tasks require finding the unique source video from a large number of real videos, while retrieval methods are typically used to find similar videos. In addition, training an effective hashing center to distinguish similar real videos is challenging. To address the above issues, we introduce a novel loss function, hash triplet loss, to capture fine-grained features with subtle differences. Extensive experiments show that our method outperforms state-of-the-art methods on multiple datasets of object removal (video inpainting), object addition (video splicing), and object swapping (face swapping), demonstrating excellent robustness and cross-dataset performance. The effectiveness of the hash triplet loss for nondifferentiable optimization problems is validated through experiments in similar video scenes.

Recent study has found out that after multiple times of recompression, the DCT coefficients of JPEG image can form an embedding domain that is robust to recompression, which is called transport channel matching (TCM) method. Because the cost function of the adaptive steganography does not consider the impact of modification on the robustness, the modified DCT coefficients of the stego image after TCM will change after recompression. To reduce the number of changed coefficients after recompression, this paper proposes a robust steganography algorithm which dynamically updates the robustness cost of every DCT coefficient. The robustness cost proposed is calculated by testing whether the modified DCT coefficient can resist recompression in every step of STC embedding process. By adding robustness cost to the distortion cost and using the framework of STC embedding algorithm to embed the message, the stego images have good performance both in robustness and security. The experimental results show that the proposed algorithm can significantly enhance the robustness of stego images, and the embedded messages could be extracted correctly at almost all cases when recompressing with a lower quality factor and recompression process is known to the user of proposed algorithm.