Per-Modality Uncertainty in Multimodal Sensor Fusion: Evidential Architecture versus Training-Time Fault Exposure
2026 10th International Artificial Intelligence and Data Processing Symposium (IDAP), Malatya, Türkiye, 5 - 06 Eylül 2026, ss.1-9, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/idap71869.2026.11727908
- Basıldığı Şehir: Malatya
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.1-9
- Karadeniz Teknik Üniversitesi Adresli: Evet
Özet
Evidential fusion architectures are often credited with graceful degradation under sensor faults, but it is unclear whether that credit belongs to the architecture or to the training procedure. We run a 2 × 4 factorial ablation, evidential UGF versus a plain late-fusion average (LF-avg) crossed with four training regimes, on two public occupancy benchmarks and 12 seeds. A dead-sensor fault is, in standardized space, identical to the modality-dropout augmentation used to train robust fusion models, so dropout-trained models face this fault in-distribution while non-augmented baselines face it out-of-distribution. This confound likely inflates prior dead-sensor robustness numbers. With training exposure held fixed, LF-avg matches or exceeds UGF’s uncertainty-fault signal in seven of eight comparisons; architecture matters mainly through the bounded-versus-unbounded influence a single miscalibrated modality can exert under the two fusion rules. Per-modality uncertainty also fails as a maintenance alert: the configuration most damaged by a dead sensor