backend/tests/test_analysis.py

from io import BytesIO

import pytest
from PIL import Image, PngImagePlugin

from app.analysis import (
    _huggingface_detector_for_model,
    _map_classifier_outputs_to_ai_probability,
    _model_inference_views,
    analyze_image_bytes,
    aggregate_verdict,
    DetectorSignal,
    extract_metadata,
    model_ensemble_signal,
)


def _png_bytes(with_marker: bool = False) -> bytes:
    image = Image.new("RGB", (96, 96), (120, 80, 200))
    output = BytesIO()
    pnginfo = PngImagePlugin.PngInfo()
    if with_marker:
        pnginfo.add_text("Software", "Stable Diffusion")
        pnginfo.add_text("prompt", "synthetic portrait generated by a diffusion model")
    image.save(output, format="PNG", pnginfo=pnginfo)
    return output.getvalue()


def test_analysis_reports_generative_metadata_marker() -> None:
    result = analyze_image_bytes(_png_bytes(with_marker=True))

    assert result["verdict"]["label"] == "likely_ai_generated"
    assert result["technical_appendix"]["metadata"]["generative_markers"]
    assert result["technical_appendix"]["hashes"]["sha256"]


def test_analysis_is_cautious_without_strong_signals() -> None:
    result = analyze_image_bytes(_png_bytes(with_marker=False))

    assert result["verdict"]["label"] in {"inconclusive", "likely_manipulated_or_deepfake"}
    assert "Image authenticity cannot be proven from pixels alone." in result["summary"]["limitations"]


def test_analysis_includes_explainable_analytical_layers() -> None:
    result = analyze_image_bytes(_png_bytes(with_marker=False))
    layers = result["analytical_layers"]

    assert len(layers) >= 8
    assert {layer["id"] for layer in layers} >= {
        "source_provenance",
        "visual_model_consensus",
        "model_transform_robustness",
        "luminance",
        "chroma",
        "edge_geometry",
        "noise_residual",
        "compression_ela",
        "frequency",
        "tile_regions",
    }
    for layer in layers:
        assert layer["question"]
        assert layer["method"]
        assert layer["conclusion"]
        assert 0 <= layer["ai_signal"] <= 1
        assert 0 <= layer["manipulation_signal"] <= 1
        assert layer["evidence"]
        assert layer["decision_role"] in {
            "primary_evidence",
            "guard_or_supporting_evidence",
            "review_context_only",
        }
        assert 0 <= layer["influence"] <= 1
        assert layer["counterfactual"]
    assert result["explainability"]["layer_ledger"]["layers"]
    assert result["explainability"]["decision_attribution"]
    assert result["explainability"]["explanation_contract"]["score_semantics"]
    assert result["explainability"]["decision_support"]["primary_drivers"]
    assert result["explainability"]["regional_evidence_map"]["grid"] == {"rows": 4, "cols": 4}

    tile_layer = next(layer for layer in layers if layer["id"] == "tile_regions")
    tile_grid = tile_layer["metrics"]["tile_grid"]
    assert len(tile_grid) == 16
    assert {tile["row"] for tile in tile_grid} == {1, 2, 3, 4}
    assert {tile["col"] for tile in tile_grid} == {1, 2, 3, 4}
    assert {tile["severity_band"] for tile in tile_grid} <= {"low", "medium", "high"}


def test_xmp_markers_are_summarized_without_raw_xml() -> None:
    image = Image.new("RGB", (96, 96), (120, 80, 200))
    output = BytesIO()
    image.save(output, format="JPEG")
    image_bytes = output.getvalue() + (
        b'<x:xmpmeta xmlns:x="adobe:ns:meta/">'
        b'<rdf:Description xmlns:photoshop="http://ns.adobe.com/photoshop/1.0/" />'
        b"</x:xmpmeta>"
    )

    metadata = extract_metadata(Image.open(BytesIO(image_bytes)), image_bytes)
    software_summary = " ".join(metadata["software_values"])

    assert metadata["editing_markers"] == ["photoshop"]
    assert "XMP metadata mentions: photoshop" in software_summary
    assert "<x:xmpmeta" not in software_summary
    assert metadata["xmp_present"] is True
    assert "raw XMP omitted" in metadata["xmp_excerpt"]
    assert "<x:xmpmeta" not in metadata["xmp_excerpt"]


def test_model_label_mapping_handles_ai_and_human_labels() -> None:
    assert _map_classifier_outputs_to_ai_probability(
        [{"label": "ai", "score": 0.91}, {"label": "hum", "score": 0.09}]
    ) == pytest.approx(0.91)
    assert _map_classifier_outputs_to_ai_probability(
        [{"label": "hum", "score": 0.82}, {"label": "ai", "score": 0.18}]
    ) == pytest.approx(0.18)


def test_model_label_mapping_sums_multiclass_ai_labels() -> None:
    assert _map_classifier_outputs_to_ai_probability(
        [
            {"label": "Real", "score": 0.23},
            {"label": "Deepfake", "score": 0.46},
            {"label": "Artificial", "score": 0.31},
        ]
    ) == pytest.approx(0.77)


def test_model_label_mapping_handles_human_artificial_labels() -> None:
    assert _map_classifier_outputs_to_ai_probability(
        [{"label": "human", "score": 0.99}, {"label": "artificial", "score": 0.01}]
    ) == pytest.approx(0.01)


def test_model_label_mapping_supports_explicit_numeric_labels() -> None:
    profile = {"ai_labels": ["label_1"], "real_labels": ["label_0"]}
    assert _map_classifier_outputs_to_ai_probability(
        [{"label": "LABEL_1", "score": 0.83}, {"label": "LABEL_0", "score": 0.17}],
        profile,
    ) == pytest.approx(0.83)


def test_broad_primary_uses_five_robustness_views() -> None:
    image = Image.new("RGB", (320, 240), (80, 110, 140))
    views = _model_inference_views(image, {"multi_view": True})

    assert [name for name, _ in views] == [
        "original",
        "center_crop_92pct",
        "jpeg_quality_85",
        "horizontal_flip",
        "social_resize_75pct",
    ]
    assert views[-1][1].size == (240, 180)


def test_model_ensemble_signal_summarizes_votes() -> None:
    signal = model_ensemble_signal(
        [
            DetectorSignal("hf:a", "ok", "model_likely_ai_generated", 0.96, None, "medium", [], 0.55),
            DetectorSignal("hf:b", "ok", "model_likely_ai_generated", 0.92, None, "medium", [], 0.55),
            DetectorSignal("hf:c", "ok", "model_likely_ai_generated", 0.90, None, "medium", [], 0.55),
        ]
    )

    assert signal.label == "ensemble_likely_ai_generated"
    assert signal.confidence == "high"
    assert signal.ai_probability == pytest.approx(0.869)


def test_model_ensemble_rejects_soft_leaning_ai_consensus() -> None:
    signal = model_ensemble_signal(
        [
            DetectorSignal("hf:a", "ok", "model_likely_ai_generated", 0.99, None, "medium", [], 0.55),
            DetectorSignal("hf:b", "ok", "model_inconclusive", 0.69, None, "medium", [], 0.55),
            DetectorSignal("hf:c", "ok", "model_inconclusive", 0.62, None, "medium", [], 0.55),
        ]
    )

    assert signal.label == "ensemble_inconclusive"
    assert signal.confidence == "low"
    assert signal.ai_probability == pytest.approx(0.6607)


def test_model_ensemble_accepts_primary_anchored_alignment() -> None:
    signal = model_ensemble_signal(
        [
            DetectorSignal(
                "hf:buildborderless/CommunityForensics-DeepfakeDet-ViT",
                "ok",
                "model_inconclusive",
                0.72,
                None,
                "low",
                [],
                1.0,
            ),
            DetectorSignal(
                "hf:Ateeqq/ai-vs-human-image-detector",
                "ok",
                "model_likely_ai_generated",
                0.999,
                None,
                "medium",
                [],
                0.3,
            ),
            DetectorSignal(
                "hf:jacoballessio/ai-image-detect-distilled",
                "ok",
                "model_inconclusive",
                0.62,
                None,
                "low",
                [],
                0.45,
            ),
        ]
    )

    assert signal.label == "ensemble_likely_ai_generated"
    assert signal.confidence == "medium"
    assert signal.ai_probability == pytest.approx(0.74)


def test_model_ensemble_is_inconclusive_when_detectors_disagree_strongly() -> None:
    signal = model_ensemble_signal(
        [
            DetectorSignal("hf:a", "ok", "model_likely_ai_generated", 0.99, None, "medium", [], 0.55),
            DetectorSignal("hf:b", "ok", "model_likely_human_or_real", 0.22, None, "medium", [], 0.55),
        ]
    )

    assert signal.label == "ensemble_inconclusive"
    assert signal.confidence == "low"


def test_aggregate_verdict_caps_uncorroborated_model_false_positive() -> None:
    detectors = [
        DetectorSignal("metadata_provenance", "ok", "inconclusive", 0.5, 0.25, "low", ["No strong metadata signal."], 0.22),
        DetectorSignal("compression_noise_forensics", "ok", "no_strong_forensic_signal", 0.38, 0.38, "low", ["Weak forensic signal."], 0.28),
        DetectorSignal("hf:a", "ok", "model_likely_ai_generated", 0.99, None, "medium", ["a"], 0.55),
        DetectorSignal("hf:b", "ok", "model_inconclusive", 0.61, None, "medium", ["b"], 0.55),
        DetectorSignal("open_source_model_ensemble", "ok", "ensemble_inconclusive", 0.8, None, "low", ["disagreement"], 0.9),
    ]
    verdict = aggregate_verdict(
        detectors,
        metadata={"generative_markers": [], "has_exif": False},
        c2pa={"claim": None},
        forensics={"artificiality_score": 0.08, "manipulation_score": 0.38},
    )

    assert verdict["label"] == "inconclusive"
    assert verdict["confidence"] == "low"
    assert verdict["ai_probability"] <= 0.64
    assert any("capped" in item for item in verdict["rationale"])


def test_aggregate_verdict_requires_independent_support_for_two_model_ai_claim() -> None:
    detectors = [
        DetectorSignal("metadata_provenance", "ok", "inconclusive", 0.5, 0.25, "low", ["No strong metadata signal."], 0.22),
        DetectorSignal("compression_noise_forensics", "ok", "no_strong_forensic_signal", 0.34, 0.29, "low", ["Weak forensic signal."], 0.28),
        DetectorSignal("hf:a", "ok", "model_likely_ai_generated", 0.99, None, "medium", ["a"], 0.55),
        DetectorSignal("hf:b", "ok", "model_likely_ai_generated", 0.88, None, "medium", ["b"], 0.55),
        DetectorSignal("open_source_model_ensemble", "ok", "ensemble_likely_ai_generated", 0.935, None, "high", ["consensus"], 0.9),
    ]
    verdict = aggregate_verdict(
        detectors,
        metadata={"generative_markers": [], "has_exif": False},
        c2pa={"claim": None},
        forensics={"artificiality_score": 0.06, "manipulation_score": 0.29},
    )

    assert verdict["label"] == "inconclusive"
    assert verdict["confidence"] == "low"
    assert verdict["ai_probability"] <= 0.55
    assert any("independent metadata" in item for item in verdict["rationale"])


def test_aggregate_verdict_allows_overwhelming_model_consensus_without_real_vote() -> None:
    detectors = [
        DetectorSignal("metadata_provenance", "ok", "inconclusive", 0.5, 0.25, "low", ["No strong metadata signal."], 0.22),
        DetectorSignal("compression_noise_forensics", "ok", "no_strong_forensic_signal", 0.36, 0.29, "low", ["Weak forensic signal."], 0.28),
        DetectorSignal("hf:a", "ok", "model_likely_ai_generated", 0.97, None, "medium", ["a"], 0.55),
        DetectorSignal("hf:b", "ok", "model_likely_ai_generated", 0.94, None, "medium", ["b"], 0.55),
        DetectorSignal("hf:c", "ok", "model_likely_ai_generated", 0.92, None, "medium", ["c"], 0.55),
        DetectorSignal("open_source_model_ensemble", "ok", "ensemble_likely_ai_generated", 0.9433, None, "high", ["consensus"], 0.9),
    ]
    verdict = aggregate_verdict(
        detectors,
        metadata={"generative_markers": [], "has_exif": False, "format": "PNG"},
        c2pa={"claim": None},
        forensics={"artificiality_score": 0.06, "manipulation_score": 0.29},
    )

    assert verdict["label"] == "likely_ai_generated"
    assert verdict["confidence"] == "medium"
    assert verdict["ai_probability"] == pytest.approx(0.72)


def test_aggregate_verdict_allows_primary_anchored_consensus() -> None:
    detectors = [
        DetectorSignal("metadata_provenance", "ok", "inconclusive", 0.5, 0.25, "low", ["none"], 0.22),
        DetectorSignal("compression_noise_forensics", "ok", "no_strong_forensic_signal", 0.36, 0.25, "low", ["weak"], 0.28),
        DetectorSignal("hf:buildborderless/CommunityForensics-DeepfakeDet-ViT", "ok", "model_inconclusive", 0.72, None, "low", ["primary"], 1.0),
        DetectorSignal("hf:Ateeqq/ai-vs-human-image-detector", "ok", "model_likely_ai_generated", 0.999, None, "medium", ["counter"], 0.3),
        DetectorSignal("hf:jacoballessio/ai-image-detect-distilled", "ok", "model_inconclusive", 0.62, None, "low", ["counter"], 0.45),
        DetectorSignal("open_source_model_ensemble", "ok", "ensemble_likely_ai_generated", 0.74, None, "medium", ["aligned"], 0.9),
    ]
    verdict = aggregate_verdict(
        detectors,
        metadata={"generative_markers": [], "has_exif": False, "format": "PNG"},
        c2pa={"claim": None},
        forensics={"artificiality_score": 0.12, "manipulation_score": 0.25, "quality": {"risk_score": 0.1}},
    )

    assert verdict["label"] == "likely_ai_generated"
    assert verdict["ai_probability"] == pytest.approx(0.72)
    assert any("primary-anchored" in item for item in verdict["rationale"])


def test_aggregate_verdict_caps_camera_like_real_photo_even_with_model_consensus() -> None:
    detectors = [
        DetectorSignal("metadata_provenance", "ok", "camera_metadata_present", 0.35, 0.25, "low", ["EXIF"], 0.22),
        DetectorSignal("compression_noise_forensics", "ok", "no_strong_forensic_signal", 0.36, 0.29, "low", ["Weak forensic signal."], 0.28),
        DetectorSignal("hf:a", "ok", "model_likely_ai_generated", 0.97, None, "medium", ["a"], 0.55),
        DetectorSignal("hf:b", "ok", "model_likely_ai_generated", 0.94, None, "medium", ["b"], 0.55),
        DetectorSignal("hf:c", "ok", "model_likely_ai_generated", 0.92, None, "medium", ["c"], 0.55),
        DetectorSignal("open_source_model_ensemble", "ok", "ensemble_likely_ai_generated", 0.9433, None, "high", ["consensus"], 0.9),
    ]
    verdict = aggregate_verdict(
        detectors,
        metadata={"generative_markers": [], "has_exif": True, "format": "JPEG"},
        c2pa={"claim": None},
        forensics={
            "artificiality_score": 0.06,
            "manipulation_score": 0.29,
            "noise": {"tile_variance_mean": 240.0, "low_noise_hint": 0.0},
            "entropy": 6.5,
        },
    )

    assert verdict["label"] != "likely_ai_generated"
    assert verdict["ai_probability"] <= 0.48
    assert any("real-photo false-positive guard" in item for item in verdict["rationale"])


def test_aggregate_verdict_uses_portrait_specialist_to_prevent_ai_false_positive() -> None:
    detectors = [
        DetectorSignal("metadata_provenance", "ok", "inconclusive", 0.5, 0.25, "low", ["No strong metadata signal."], 0.22),
        DetectorSignal("compression_noise_forensics", "ok", "no_strong_forensic_signal", 0.36, 0.29, "low", ["Weak forensic signal."], 0.28),
        DetectorSignal("hf:Ateeqq/ai-vs-human-image-detector", "ok", "model_likely_ai_generated", 0.99, None, "medium", ["a"], 0.72),
        DetectorSignal("hf:dima806/ai_vs_real_image_detection", "ok", "model_likely_ai_generated", 0.87, None, "medium", ["b"], 0.35),
        DetectorSignal("hf:jacoballessio/ai-image-detect-distilled", "ok", "model_likely_human_or_real", 0.26, None, "medium", ["c"], 0.70),
        DetectorSignal("hf:SadraCoding/SDXL-Deepfake-Detector", "ok", "model_likely_human_or_real", 0.0, None, "medium", ["portrait"], 0.65),
        DetectorSignal("open_source_model_ensemble", "ok", "ensemble_inconclusive", 0.5, None, "low", ["tie"], 0.9),
    ]
    verdict = aggregate_verdict(
        detectors,
        metadata={"generative_markers": [], "has_exif": False},
        c2pa={"claim": None},
        forensics={
            "artificiality_score": 0.06,
            "manipulation_score": 0.38,
            "quality": {"risk_score": 0.28},
        },
    )

    assert verdict["label"] == "inconclusive"
    assert verdict["confidence"] == "low"
    assert verdict["ai_probability"] <= 0.52


def test_portrait_specialist_is_skipped_when_portrait_gate_is_low() -> None:
    signal = _huggingface_detector_for_model(
        Image.new("RGB", (256, 256), (30, 40, 50)),
        "SadraCoding/SDXL-Deepfake-Detector",
        {"portrait_score": 0.42, "evidence": ["low portrait score"]},
    )

    assert signal.status == "unavailable"
    assert signal.label == "portrait_gate_not_met"
    assert signal.ai_probability is None