Twitter-COMMs: Detecting Climate, COVID, and Military Multimodal Misinformation

Abstract

Detecting out-of-context media, such as "miscaptioned" images on Twitter, often requires detecting inconsistencies between the two modalities. This paper describes our approach to the Image-Text Inconsistency Detection challenge of the DARPA Semantic Forensics (SemaFor) Program. First, we collect Twitter-COMMs, a large-scale multimodal dataset with 884k tweets relevant to the topics of Climate Change, COVID-19, and Military Vehicles. We train our approach, based on the state-of-the-art CLIP model, leveraging automatically generated random and hard negatives. Our method is then tested on a hidden human-generated evaluation set. We achieve the best result on the program leaderboard, with 11% detection improvement in a high precision regime over a zero-shot CLIP baseline. * Denotes equal contribution. 1 https://www.darpa.mil/program/semantic-forensics

Synthetic Misinformers: Generating and Combating Multimodal Misinformation

“…The first MMD approaches mostly relied on convolutional neural networks pre-trained on ImageNet to extract features from images (namely VGG-19 [13,26] and ResNet50 [5,22]) and word embeddings to extract features from captions (namely word2vec [13,26] and fastText [22]). More recent approaches have resorted to large-scale multimodal and cross-modal models, namely CLIP [7,12,20], VisualBERT [20] and VinVL [12] to extract both their visual and textual features. In the aforementioned works, CLIP [24] tended to outperform other cross-modal methods (VinVL and VisualBERT) for MMD [7,12,20].…”

“…OOC involves pairing an image with an incongruous caption [5] while NEI involves manipulating the named entities in otherwise truthful captions [22]. Previous works have relied on random sampling [5,13] or feature-informed sampling methods [7,20] for generating OOC and in-cluster random sampling [26] or rule-based random sampling [22] for generating NEI. Examples of generated OOC and NEI misinformation can be seen in Fig.…”

“…One common form of multimodal misinformation involves de-contextualization; a legitimate image being paired with an out-of-context caption creating a deceptive impression. Consequently, researchers have used random sampling ( [5,13]) and feature-informed sampling methods ( [7,20]) for generating OOCs. The MAIM dataset was created by randomly sampling image-text pairs from image-caption pairs collected from Flickr [13].…”

“…To this end, Luo et al [20] created the NewsCLIPings dataset by utilizing the large cross-modal CLIP model [24] along with scene-learning and person matching models in order to generate hard negative samples. Similarly, the Twitter-COMMs dataset was created by combining and applying CLIP-based sampling (to generate hard negatives) and in-topic random sampling (to resolve class imbalance) on data collected from Twitter, related to three topics: climate, COVID, and military vehicles [7].…”

“…The above works either provide internal ablation [7,20,22,26] or comparison with simple baselines [26] and do not compare their methods with other Synthetic Misinformers. Moreover, studies that used some of the above datasets have mostly focused on incremental methodological improvements [2], such as the integration of sentiment [4] or evidence [1] in MMD models.…”