Fakebench Eval

fakebench-eval

FakeBench: Probing Explainable Fake Image Detection via Large Multimodal Models — Yixuan Li et al. (2024) (arXiv:2404.13306, 2024)

What this evaluates

Evaluates large multimodal models on explainable fake image detection across closed-ended classification and open-ended reasoning tasks. It probes the models' ability to accurately classify image authenticity and generate evidence-based, interpretable justifications using visual and textual forensic cues.

Datasets

• FakeBench — total ?; splits: test (-1); repo https://github.com/Yixuan423/FakeBench

Metrics

• Accuracy (ACC) (primary) — range: [0, 1]
  • Ratio of correct responses to the total number of questions in the closed-ended FakeClass subset.
• BLEU-1 / BLEU-2 / ROUGE-L / Sentence Similarity — range: [0, 1]
  • Standard automatic text similarity metrics computed between model-generated responses and golden clues under a narrative setting.
• LLM-as-judge (Completeness, Preciseness, Relevance) — range: [0, 2]
  • GPT-4 rates responses on a scale of {0,1,2} for completeness, preciseness, and relevance. Scores are averaged over 5 repeated rating rounds per image, then macro-averaged across the dataset.

Input / output format

Input: An image paired with a question (closed-ended for classification, or open-ended for reasoning/clues/QA).

Output: For closed-ended: a binary or categorical classification response. For open-ended: a natural language explanation or answer.

Scoring recipe

def compute_metrics(predictions, gold):
    # Accuracy (FakeClass)
    acc = sum(1 for p, g in zip(predictions, gold) if p == g) / len(gold)
    
    # LLM-as-judge (FakeClue/FakeQA)
    # For each image, run GPT-4 judge 5 times to get scores for Comp, Prec, Rele in {0,1,2}
    # Average the 5 rounds per image, then macro-average across all images
    avg_scores = []
    for img_preds, img_gold in zip(predictions, gold):
        round_scores = [judge_gpt4(img_preds, img_gold) for _ in range(5)]
        avg_scores.append(mean(round_scores))
    overall_score = mean(avg_scores)
    return acc, overall_score

Common pitfalls

• LLM-as-a-judge scores are averaged over 5 repeated rating rounds per image to ensure reliability, which is often overlooked in reproduction.
• Specialized unimodal models are evaluated under a specific generalization training setting (360K ProGAN fake + 360K LSUN real images), not zero-shot like the LMMs.
• Automatic text similarity metrics (BLEU/ROUGE) are used alongside LLM-as-judge, but the LLM judge uses a proprietary GPT-4 with specific prompts provided only in the supplementary material.

Evidence (verbatim from paper)

For FakeClass, which is composed of closed-ended questions, we compute the correctness of the model's responses in comparison to the standard responses. The accuracy (ACC) is utilized as the metric to assess LMMs in fake image detection, which is computed by the ratio of correct responses to the total number of questions in FakeClass. For FakeClue and FakeQA encompassing open-ended questions, we compute text similarities under the narrative setting between model responses and the golden clues inspired by recent NLP studies [92], [93]. Aligned with previous work [24], [73], [74], [94]–[97], we report automatic metrics (BLEU (B.)-1/2 [98], ROUGE (R.)-L [99], Sentence Similarity (Sim.) [100]), and the LLM-as-a-judge method [93] to evaluate model responses. In particular, the LLM-as-a-judge evaluation focuses on three dimensions: (1) Completeness (Comp.): More information aligning with the golden clues is preferred. (2) Preciseness (Prec.): Information that conflicts with the golden clues will be penalized. (3) Relevance (Rele.): More information should be closely related to image authenticity. For each image, the GPT-assisted measures select one value from {0,1,2}, indicating the low

Citation

@misc{li2024fakebench,
  title={FakeBench: Probing Explainable Fake Image Detection via Large Multimodal Models},
  author={Yixuan Li et al. (2024)},
  year={2024},
  note={arXiv:2404.13306}
}

• arXiv: 2404.13306