Study finds when trusting graph relations hurts detectors
Original title: When Is a Relation Worth Trusting? Manipulation Costs, Exposure, and the Trust Trap in Multi-Relational Graph Learning
Clean-trained graph detectors fell below graph-free counterparts in 58 of 90 settings on six graphs.
A paper analyzes when multi-relational graph detectors should ignore a relation because an adversary faces different costs to manipulate each relation. It proves a sharp case where degree gating beats selection or reweighting if part of the population is isolated and the attacker cannot open the gate. Exposure per unit cost selected the relation whose hardening raised the trap threshold most in 15 of 16 cases. Without cost information, the authors say the safe choice is to abstain.
Unconfirmed: Cost-aware improvements with many relations are reported so far only on synthetic graphs.
- Validated gating gains occurred only on relations meeting the stated conditions.
- Informativeness, cheapness, or equal edit costs succeeded in at most 7 cases.
- The selection protocol was never significantly worse than ignoring the graph.
- With at most four relations, the protocol mostly abstained.
- Cost-aware gains with many relations were shown so far on synthetic graphs.
From the original article
1 min read at arxiv.orgMulti-relational graph detectors learn from clean data how much to trust each relation, but an adversary pays a different price to manipulate each relation and each action; taking such costs as given, we ask when using a relation is worse than ignoring it.
By Vincess Dongmo · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.