Robust Image Hashing for Detecting Small Tampering Using a Hyperrectangular Region (10:50 AM – 11:10 AM)
Toshiki Itagaki (Sony Semiconductor Solutions Corporation), Yuki Funabiki (Sony Corporation) and Toru Akishita (Sony Semiconductor Solutions Corporation) – Virtual presentation
In this paper, we propose a robust image hashing method that enables detecting small tampering. Existing hashing methods are too robust, and the trade-off relation between the robustness and the sensitivity to visual content changes needs to be improved to detect small tampering. Though the adaptive thresholding method can improve the trade-off, there’s more room to improve and it requires tampered image derived from the original, which limits its applications. To overcome these two drawbacks, we introduce a new concept of a hyperrectangular region in multi-dimensional hash space, which is determined at the timing of hash generation as the region that covers a hash cluster by using the maximum and the minimum of the cluster per each hash axis. We evaluate our method and the existing methods. Our method improves the trade-off, which achieves 0.9428 as AUC (Area Under the Curve) for detecting tampering that occupies about 0.1% area of the image in the presence of JPEG compression and reducing the size as content-preserving operations. Furthermore, our method does not require tampered image derived from the original, which differs from the existing method.
Scalable Fact-checking with Human-in-the-Loop (11:10 AM – 11:30 AM)
Jing Yang (University of Campinas), Didier Augusto Vega-Oliveros (University of Campinas), Taís Seibt (Universidade do Vale do Rio dos Sino) and Anderson Rocha (UNICAMP) – Virtual presentation
Researchers have been investigating automated solutions for fact-checking in various fronts. However, current approaches often overlook the fact that information released every day is escalating, and a large amount of them overlap. Intending to accelerate fact-checking, we bridge this gap by proposing a new pipeline — grouping similar messages and summarizing them into aggregated claims. Specifically, we first clean a set of social media posts (e.g., tweets) and build a graph of all posts based on their semantics; Then, we perform two clustering methods to group the messages for further claim summarization. We evaluate the summaries both quantitatively with ROUGE scores and qualitatively with human evaluation. We also generate a graph of summaries to verify that there is no significant overlap among them. The results reduced 28,818 original messages to 700 summary claims, showing the potential to speed up the fact-checking process by organizing and selecting representative claims from massive disorganized and redundant messages.
Source Attribution of Online News Images by Compression Analysis (11:30 AM – 11:50 AM)
Michael Albright (Kitware Inc.), Nitesh Menon (Kitware Inc.), Kristy Roschke (Arizona State University) and Arslan Basharat (Kitware Inc.) – Virtual presentation
The rapid increase in the amount of online disinformation warrants new and robust digital forensics methods for validating purported sources of multimodal news articles. We conducted a survey of news photojournalists for insights into their workflows. A high percentage (91%) of respondents reported standardized photo publishing procedures, which we hypothesize facilitates source verification. In this work, we demonstrate that the online news sites leave predictable and discernible patterns in the compression settings of the images they publish. We propose novel, simple, and very efficient algorithms to analyze the image compression profiles for news source verification and identification. We evaluate the algorithms’ effectiveness through extensive experiments on a newly-released dataset of over 64K images from over 34K articles collected from 30 news sites. The image compression features are modeled by Naive Bayes variants or XGBoost classifiers for source attribution and verification. For these news sources we are able to achieve very strong performance with the proposed algorithms resulting in 0.92 – 0.94 average AUC for source verification under a closed set scenario, and compelling open set generalization with only 0.0 – 0.04 reduction in the average AUC.
Time Scaling Detection and Estimation in Audio Recordings (11:50 AM – 12:10 PM)
Michele Pilia (Politecnico di Milano), Sara Mandelli (Politecnico di Milano), Paolo Bestagini (Politecnico di Milano) and Stefano Tubaro (Politecnico di Milano, Italy) – Virtual presentation
The widespread diffusion of user friendly editing software for audio signals has made audio tampering extremely accessible to anyone. Therefore, it is increasingly necessary to develop forensic methodologies aiming at verifying if a given audio content has been digitally manipulated or not. Among the multiple available audio editing techniques, a very common one is time scaling, i.e., altering the temporal evolution of an audio signal without affecting any pitch component. For instance, this can be used to slow-down or speed-up speech recordings, thus enabling the creation of natural sounding fake speech compositions. In this work, we propose to blindly detect and estimate the time scaling applied to an audio signal. To expose time scaling, we leverage a Convolutional Neural Network that analyzes the Log-Mel Spectrogram and the phase of the Short Time Fourier Transform of the input audio signal. The proposed technique is tested on different audio datasets, considering various time scaling implementations and challenging cross test scenarios.
