When building intelligent systems that perceive the world through multiple lenses—whether they are autonomous vehicles balancing LiDAR and cameras, or clinical models combining vital signs with doctor notes—disagreements between data sources are inevitable. A camera might see a clear road ahead, while a radar sensor detects an obstacle.
Traditional statistical frameworks often struggle when inputs directly contradict one another. If we simply average the signals or apply rigid probabilistic updates, the system can arrive at a consensus that is not only incorrect but dangerously overconfident. The field of conflict-sensitive evidence fusion focuses on a core question: How can a system mathematically reason through extreme contradictions without destroying the underlying truth?
The Core Paradox: Why Standard Fusion Fails
To understand why specialized conflict-resolution mechanisms are necessary, we must look at how classical frameworks handle uncertainty. In standard Bayesian probability, a system must distribute a total probability mass of exactly $1.0$ across all possible mutually exclusive outcomes. As noted by Sven Richter et al. in Fusion of range measurements and semantic estimates in an evidential framework, this requirement makes it difficult to explicitly represent complete ignorance or a fundamental conflict between sensors; the math forces the system to choose or average them out.
To bypass this limitation, researchers frequently turn to Dempster-Shafer (D-S) evidence theory, an early foundation for which was laid by Thomas D. Garvey et al. in An Inference Technique for Integrating Knowledge from Disparate Sources. D-S theory allows sensors to assign "basic probability assignments" (BPA) not just to single outcomes, but to sets of outcomes (e.g., "either a car or a pedestrian").
However, standard D-S theory contains a critical vulnerability known as Zadeh’s Paradox. When two sources highly conflict, the standard Dempster’s rule of combination normalizes the joint belief by dividing by a factor of $1 - K$, where $K$ represents the metric of conflict. If the conflict is near-total ($K \to 1$), a tiny, accidental sliver of shared agreement can be mathematically inflated into absolute certainty.
As Didier Dubois et al. demonstrate in Combination and Propagation of Uncertainty with Belief Functions - A Reexamination, this aggressive normalization produces counter-intuitive, overly certain results that are highly sensitive to minor fluctuations in input. Under these conditions, Ben P. Wise et al. show in A Framework for Comparing Uncertain Inference Systems to Probability that standard evidence theory can actually perform worse than random chance.
Mechanisms for Resolving Conflict
To prevent conflicting evidence from poisoning the final output, modern research focuses on three primary strategies: redistributing the conflict mass, pre-weighting evidence via entropy, and adaptive rejection.
1. Redistributing Conflict (DSmT and Transferable Belief)
Instead of forcing a consensus through normalization, alternative fusion rules preserve or redistribute the conflict mass. A prominent example is Dezert-Smarandache Theory (DSmT). As explored by Xinhan Huang et al. in Robot Map Building From Sonar Sensors and DSmT, DSmT relaxes the assumption of strict mutual exclusivity, allowing the system to explicitly track overlapping or contradictory hypotheses.
By avoiding standard normalization, the system can maintain a cautious stance when sensors fundamentally disagree. For instance, Haibin Yu et al. applied this framework in A multisource fusion framework driven by user-defined knowledge for egocentric activity recognition to successfully adjudicate scenarios where visual data and body-worn inertial sensors provided completely opposite activity cues.
Similarly, the Transferable Belief Model proposed by Philippe Smets et al. in Default reasoning and the transferable belief model separates the concept of holding beliefs from the act of making a decision, allowing conflict to be held in an "open-world" state until a decision rule forces a resolution.
2. Pre-weighting and Belief Entropy
A second approach calculates the reliability of each source before combining them. If a source is highly uncertain or wildly out of alignment with the rest of the pool, its influence is down-weighted.
Fuyuan Xiao 0001 develops this intuition in Combine Conflicting Evidence Based on the Belief Entropy and IOWA Operator, using belief entropy to measure the internal uncertainty of each source alongside an Induced Ordered Weighted Averaging (IOWA) operator. This allows the system to dynamically adjust the authority of a sensor based on how chaotic its readings are relative to the group.
To handle open-world scenarios where a sensor might encounter an entirely unknown class, Dongdong Wu et al. introduce a method in A new approach for generation of generalized basic probability assignment in the evidence theory that leverages triangular fuzzy numbers to construct generalized BPAs, minimizing information loss when contradictions arise from unexpected environmental states.
[Sensor A] ---\
--> [Calculate Belief Entropy / Distance] --> [Assign Weights] --> [Conflict-Free Fusion]
[Sensor B] ---/
3. Trust and Adaptive Rejection in Deep Learning
In modern multi-view deep learning, different neural network branches act as distinct "views" or experts. When these views provide highly contradictory predictions, a single corrupted view can easily mislead the entire model.
To combat this, Jueqing Lu et al. in Navigating Conflicting Views: Harnessing Trust for Learning and Wei Liu et al. in Enhancing Multi-View Classification Reliability with Adaptive Rejection introduce mechanisms that treat fusion as a group decision-making problem. By modeling outputs as formal opinions, these networks can dynamically identify a rogue, untrustworthy view and trigger an adaptive rejection mechanism, choosing to ignore the conflicting input entirely rather than attempting to reconcile it.
Real-World Applications
The mathematical necessity of handling conflicting evidence spans across several high-stakes domains:
- Autonomous Driving: Vehicles must continuously reconcile real-time sensor observations with high-level map priors. As highlighted by Fabian Immel et al. in SDTagNet: Leveraging Text-Annotated Navigation Maps for Online HD Map Construction, deciding when to trust a noisy sensor over a potentially outdated map prior requires robust conflict reasoning. To implement this safely, Frank Bieder et al. utilize data-driven reliability coefficients in Sensor Data Fusion in Top-View Grid Maps using Evidential Reasoning with Advanced Conflict Resolution to resolve direct contradictions between LiDAR and stereo cameras on occupancy grid maps.
- Clinical Diagnostics: In intensive care units, models often encounter conflicting signals, such as stable vital signs paired with a nurse's qualitative note describing a sudden patient decline. Yucheng Ruan et al. demonstrate in Towards Accurate and Reliable ICU Outcome Prediction: A Multimodal Learning Framework Based on Belief Function Theory using Structured EHRs and Free-Text Notes that using belief function theory to preserve this conflict—rather than averaging it out—leads to significantly more reliable mortality forecasts.
- Cybersecurity and Infrastructure: In Cyber-Physical Systems, a compromised node might execute a Time Delay Attack, broadcasting false timing information to trigger a failure. Yang Li demonstrates in Secure Combination of Untrusted Time information Based on Optimized Dempster-Shafer Theory that by "clipping" the probability assignments of highly conflicting, untrusted sources, the system can isolate malicious actors and maintain picosecond-level stability.
By shifting from rigid probabilistic aggregation to conflict-aware evidential reasoning, these frameworks ensure that when our sensors disagree, our systems do not blindly leap to the wrong conclusion.
Key Foundational Work
A Distributed Control System for the CMU Rover (1983) — contributed 2 concept elements to this direction.
An Inference Technique for Integrating Knowledge from Disparate Sources (1981) — contributed 1 concept elements to this direction.