Stefan Smeu

Stefan Smeu, Elena Burceanu, Emanuela Haller et al. · mila

Novelty detection seeks to identify samples deviating from a known distribution, yet data shifts in a multitude of ways, and only a few consist of relevant changes. Aligned with out-of-distribution generalization literature, we advocate for a formal distinction between task-relevant semantic or content changes and irrelevant style changes. This distinction forms the basis for robust novelty detection, emphasizing the identification of semantic changes resilient to style distributional shifts. To this end, we introduce Stylist, a method that utilizes pretrained large-scale model representations to selectively discard environment-biased features. By computing per-feature scores based on feature distribution distances between environments, Stylist effectively eliminates features responsible for spurious correlations, enhancing novelty detection performance. Evaluations on adapted domain generalization datasets and a synthetic dataset demonstrate Stylist's efficacy in improving novelty detection across diverse datasets with stylistic and content shifts. The code is available at https://github.com/bit-ml/Stylist.

Stefan Smeu, Elena Burceanu, Andrei Liviu Nicolicioiu et al. · mila

We introduce a formalization and benchmark for the unsupervised anomaly detection task in the distribution-shift scenario. Our work builds upon the iWildCam dataset, and, to the best of our knowledge, we are the first to propose such an approach for visual data. We empirically validate that environment-aware methods perform better in such cases when compared with the basic Empirical Risk Minimization (ERM). We next propose an extension for generating positive samples for contrastive methods that considers the environment labels when training, improving the ERM baseline score by 8.7%.

Stefan Smeu, Elena Burceanu, Emanuela Haller et al. · mila

We tackle the problem of robust novelty detection, where we aim to detect novelties in terms of semantic content while being invariant to changes in other, irrelevant factors. Specifically, we operate in a setup with multiple environments, where we determine the set of features that are associated more with the environments, rather than to the content relevant for the task. Thus, we propose a method that starts with a pretrained embedding and a multi-env setup and manages to rank the features based on their environment-focus. First, we compute a per-feature score based on the feature distribution variance between envs. Next, we show that by dropping the highly scored ones, we manage to remove spurious correlations and improve the overall performance by up to 6%, both in covariance and sub-population shift cases, both for a real and a synthetic benchmark, that we introduce for this task.

Stefan Smeu, Dragos-Alexandru Boldisor, Dan Oneata et al.

Good datasets are essential for developing and benchmarking any machine learning system. Their importance is even more extreme for safety critical applications such as deepfake detection - the focus of this paper. Here we reveal that two of the most widely used audio-video deepfake datasets suffer from a previously unidentified spurious feature: the leading silence. Fake videos start with a very brief moment of silence and based on this feature alone, we can separate the real and fake samples almost perfectly. As such, previous audio-only and audio-video models exploit the presence of silence in the fake videos and consequently perform worse when the leading silence is removed. To circumvent latching on such unwanted artifact and possibly other unrevealed ones we propose a shift from supervised to unsupervised learning by training models exclusively on real data. We show that by aligning self-supervised audio-video representations we remove the risk of relying on dataset-specific biases and improve robustness in deepfake detection.

Dragos-Alexandru Boldisor, Stefan Smeu, Dan Oneata et al.

Self-supervised representations excel at many vision and speech tasks, but their potential for audio-visual deepfake detection remains underexplored. Unlike prior work that uses these features in isolation or buried within complex architectures, we systematically evaluate them across modalities (audio, video, multimodal) and domains (lip movements, generic visual content). We assess three key dimensions: detection effectiveness, interpretability of encoded information, and cross-modal complementarity. We find that most self-supervised features capture deepfake-relevant information, and that this information is complementary. Moreover, models primarily attend to semantically meaningful regions rather than spurious artifacts. Yet none generalize reliably across datasets. This generalization failure likely stems from dataset characteristics, not from the features themselves latching onto superficial patterns. These results expose both the promise and fundamental challenges of self-supervised representations for deepfake detection: while they learn meaningful patterns, achieving robust cross-domain performance remains elusive.