2nd/10_Wiki/Topics/AI_and_ML/Out-of-distribution-Detection.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-out-of-distribution-detection	Out-of-Distribution Detection	10_Wiki/Topics	verified	self	OOD-Detection OOD Anomaly-Detection-NN Novelty-Detection	none	A	0.9	applied	ood safety uncertainty foundation-models anomaly-detection		2026-05-10	pending	language framework python pytorch

Out-of-Distribution Detection

매 한 줄

"매 model 이 본 적 없는 input 의 거부". OOD detection 은 inference 시 input 이 training distribution 밖인지 판정하여 silent failure 를 막는 safety layer. 매 2026 의 표준은 foundation-model embedding 위 의 KNN / Mahalanobis 또는 logit-energy score, classical ODIN 은 baseline.

매 핵심

매 score family

1. Softmax baseline (MSP): max softmax probability — weak baseline.
2. ODIN (Liang 2018): temperature scaling + input gradient perturbation.
3. Energy (Liu 2020): -T * logsumexp(logits / T), free, strong.
4. Mahalanobis (Lee 2018): class-conditional Gaussian on penultimate features.
5. KNN (Sun 2022): k-NN distance in feature space — 매 simple, robust.
6. DOSE / VIM (2022-2024): residual + logit hybrid.
7. Foundation-model OOD (CLIP, DINOv2 features + KNN) — 2026 SOTA.

매 evaluation

• AUROC, FPR@95TPR, AUPR.
• ID = CIFAR-10/ImageNet, OOD = SVHN, Textures, iNaturalist, Places, OpenOOD bench.
• near-OOD (CIFAR10 vs CIFAR100) 가 매 어려운 case.

매 응용

1. autonomous driving 의 unknown object reject.
2. medical imaging 의 unsupported modality flag.
3. LLM 의 jailbreak / off-distribution prompt detection.
4. fraud detection 의 novel attack pattern.

💻 패턴

Energy score (Liu 2020)

import torch, torch.nn.functional as F

@torch.no_grad()
def energy_score(model, x, T=1.0):
    logits = model(x)
    # higher energy = OOD
    return -T * torch.logsumexp(logits / T, dim=-1)

MSP baseline

@torch.no_grad()
def msp(model, x):
    return -F.softmax(model(x), dim=-1).max(-1).values

Mahalanobis on features

@torch.no_grad()
def fit_mahalanobis(features, labels, num_classes):
    means = []
    for c in range(num_classes):
        means.append(features[labels == c].mean(0))
    means = torch.stack(means)
    centered = features - means[labels]
    cov = centered.T @ centered / len(features)
    inv = torch.linalg.pinv(cov)
    return means, inv

def maha_score(f, means, inv):
    diffs = f.unsqueeze(1) - means    # [N, C, D]
    d2 = torch.einsum("ncd,de,nce->nc", diffs, inv, diffs)
    return d2.min(-1).values          # smallest distance to any class

KNN OOD (Sun 2022)

import torch, torch.nn.functional as F

class KNNOOD:
    def __init__(self, k=50):
        self.k = k
    def fit(self, train_feats):
        self.bank = F.normalize(train_feats, dim=-1)
    def score(self, feats):
        f = F.normalize(feats, dim=-1)
        sim = f @ self.bank.T          # cosine
        topk = sim.topk(self.k, dim=-1).values
        return -topk[:, -1]            # negative k-th similarity → OOD score

ODIN

def odin_score(model, x, T=1000, eps=0.0014):
    x = x.clone().detach().requires_grad_(True)
    logits = model(x) / T
    p = F.softmax(logits, dim=-1).max(-1).values
    p.sum().backward()
    x_adv = x - eps * x.grad.sign()
    with torch.no_grad():
        return F.softmax(model(x_adv) / T, dim=-1).max(-1).values

Foundation-model OOD (DINOv2 + KNN)

import torch
dino = torch.hub.load("facebookresearch/dinov2", "dinov2_vitb14").eval().cuda()

@torch.no_grad()
def feats(x):
    return dino(x)   # [B, 768]

knn = KNNOOD(k=50)
knn.fit(feats(train_loader_id))
ood_scores = knn.score(feats(test_batch))

LLM OOD via embedding

from sentence_transformers import SentenceTransformer
emb = SentenceTransformer("BAAI/bge-large-en-v1.5")
id_bank = emb.encode(in_dist_prompts, normalize_embeddings=True)

def prompt_ood(prompt, k=20):
    q = emb.encode([prompt], normalize_embeddings=True)
    sims = (q @ id_bank.T)[0]
    return -sims.topk(k).values.min()

Threshold calibration (FPR@95TPR)

import numpy as np
def threshold_at_tpr(scores_id, tpr=0.95):
    return np.quantile(scores_id, 1 - tpr)

매 결정 기준

상황	Method
매 simple, 즉시	Energy
매 best AUROC	KNN on foundation features
매 access to features only	Mahalanobis
매 CV with strong backbone	DINOv2 + KNN
매 LLM input filter	embedding KNN + threshold
매 production, low-latency	Energy or MSP

기본값: foundation embedding + KNN (k=50).

🔗 Graph

• 부모: Anomaly-Detection
• 변형: KNN
• 응용: Active-Learning

🤖 LLM 활용

언제: 매 high-stakes deployment, jailbreak filter, novel-prompt routing. 언제 X: 매 closed-world benchmark — distribution 가 fixed 인 경우 overhead.

❌ 안티패턴

• MSP only: 매 over-confident network 에서 거의 무력.
• Train OOD detector on test OOD set: leakage, false confidence.
• Threshold from training scores: ID validation set 에서 calibrate.
• Ignore near-OOD: far-OOD AUROC 99% 인데 near-OOD 60% 인 흔한 함정.
• Foundation-model embedding mismatch: ImageNet-pretrained 으로 medical OOD detect.

🧪 검증 / 중복

• Verified (OpenOOD benchmark 2024, Sun 2022 KNN, Liu 2020 Energy).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — Energy/Maha/KNN + foundation-model OOD