Unverified50% confidenceFactExact time
In distillation attacks, if the API exposes logits or temperature-scaled softmax outputs, attackers can train a substitute closer to the original model with fewer queries.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedReducing the amount of information returned by the API (such as returning only the final text result without exposing logits distributions, top-k probabilities, or token-level confidence) can significantly increase the difficulty of distillation attacks, because the supervision signal available to the attacker becomes sparser.73% similarUnverifiedRate limiting at the API layer increases the attacker's data collection cost by setting a maximum number of queries per unit time, greatly reducing the economic viability of distillation attacks.70% similarUnverifiedMetrics-driven optimization carries risks: agents targeting reductions in React commits or V8 call counts tend to cut network requests that don't affect interactivity, leading to overly aggressive caching with insufficient invalidation69% similarUnverifiedMembership inference attacks can detect data contamination via perplexity: if a model's perplexity for a given code segment is significantly lower than for code of equivalent complexity, this often indicates the segment appeared in the training set.67% similarUnverifiedGreedy decoding at temperature zero further suppresses the probability of a model outputting ESCALATE, because models tend to select the answer pattern that appeared most frequently during training, and escalation is inherently rare in instruction fine-tuning data.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1013964API
curl https://kongchang.com/api/v1/knowledge/claims/1013964MCP
get_claim(id=1013964)