df2 / src /tools /forensic /cfa_tools.py
Mustafa Akcanca
Add CFA
5196d55
"""
Popescu–Farid CFA Consistency Analyzer for the agent.
This tool analyzes Color Filter Array (CFA) demosaicing artifacts to detect
inconsistencies within an image. It is designed for SPLICE DETECTION and
SOURCE CONSISTENCY analysis, NOT for whole-image authenticity classification.
Scientific basis:
- Real camera images have CFA interpolation artifacts from Bayer demosaicing
- Spliced regions from different sources (AI, screenshots, different cameras)
may have different or absent CFA patterns
- By analyzing the DISTRIBUTION of CFA metrics across windows, we can identify
regions that are inconsistent with the rest of the image
What this tool DOES:
- Detects CFA pattern consistency across image regions
- Identifies outlier windows that differ from the image baseline
- Provides distribution analysis (unimodal vs bimodal)
What this tool does NOT do:
- Classify whole images as "authentic" or "fake"
- Work reliably on heavily compressed images
- Detect AI-generated images (use TruFor for that)
Supports two modes:
- analyze: run CFA consistency analysis on a single image
- calibrate: optional; build reference thresholds from a set of camera images
"""
from __future__ import annotations
import json
import sys
from pathlib import Path
from typing import Any, Dict, List, Sequence, Tuple
import numpy as np
# Ensure repo root (which contains example_tools) is on sys.path so we can load
# example_tools/cfa.py as a namespace package.
ROOT = Path(__file__).resolve().parents[3]
if str(ROOT) not in sys.path:
sys.path.append(str(ROOT))
try:
from example_tools import cfa # type: ignore
except Exception as exc: # pragma: no cover - defensive import guard
raise ImportError(
"Unable to import example_tools.cfa. Ensure repository root is on sys.path."
) from exc
DEFAULT_PATTERN = "RGGB"
DEFAULT_WINDOW = 256
DEFAULT_TOP_K = 5
DEFAULT_OUTLIER_ZSCORE = 2.0 # Windows beyond this z-score are outliers
def _parse_request(input_str: str) -> Dict[str, Any]:
"""Parse JSON or treat input_str as image_path for analyze mode."""
try:
data = json.loads(input_str)
if isinstance(data, dict):
return data
if isinstance(data, str):
return {"mode": "analyze", "image_path": data}
except Exception:
pass
return {"mode": "analyze", "image_path": input_str}
def _compute_stats(values: Sequence[float]) -> Dict[str, float]:
"""Compute basic statistics for a list of values."""
arr = np.asarray(values, dtype=np.float64)
if arr.size == 0:
return {"min": 0.0, "max": 0.0, "mean": 0.0, "median": 0.0, "std": 0.0}
return {
"min": float(np.min(arr)),
"max": float(np.max(arr)),
"mean": float(np.mean(arr)),
"median": float(np.median(arr)),
"std": float(np.std(arr)),
}
def _detect_bimodality(values: Sequence[float]) -> Dict[str, Any]:
"""
Detect if the distribution of values is bimodal using Hartigan's dip test
approximation and coefficient of bimodality.
Returns:
Dictionary with bimodality analysis results
"""
arr = np.asarray(values, dtype=np.float64)
if arr.size < 10:
return {
"is_bimodal": False,
"bimodality_coefficient": 0.0,
"distribution_type": "insufficient_data",
"note": "Need at least 10 windows for distribution analysis",
}
# Compute bimodality coefficient: BC = (skewness^2 + 1) / kurtosis
# BC > 0.555 suggests bimodality (Pfister et al., 2013)
mean = np.mean(arr)
std = np.std(arr)
if std < 1e-10:
return {
"is_bimodal": False,
"bimodality_coefficient": 0.0,
"distribution_type": "constant",
"note": "All values are nearly identical",
}
normalized = (arr - mean) / std
skewness = float(np.mean(normalized ** 3))
kurtosis = float(np.mean(normalized ** 4))
# Excess kurtosis adjustment (Fisher's definition)
excess_kurtosis = kurtosis - 3.0
# Bimodality coefficient
# For a uniform distribution: BC ≈ 0.555
# For a bimodal distribution: BC > 0.555
bc = (skewness ** 2 + 1) / (kurtosis + 3 * ((arr.size - 1) ** 2) / ((arr.size - 2) * (arr.size - 3)))
bc = float(bc)
# Also check coefficient of variation (CV) - high CV suggests mixed sources
cv = std / mean if mean > 0 else 0.0
# Determine distribution type
if bc > 0.6:
dist_type = "bimodal"
is_bimodal = True
elif bc > 0.5:
dist_type = "possibly_bimodal"
is_bimodal = False
elif cv > 0.3:
dist_type = "high_variance"
is_bimodal = False
else:
dist_type = "unimodal"
is_bimodal = False
return {
"is_bimodal": is_bimodal,
"bimodality_coefficient": bc,
"coefficient_of_variation": float(cv),
"skewness": skewness,
"excess_kurtosis": excess_kurtosis,
"distribution_type": dist_type,
}
def _find_outliers(
values: Sequence[float],
positions: Sequence[Tuple[int, int, int, int]],
z_threshold: float = DEFAULT_OUTLIER_ZSCORE,
) -> Tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
"""
Find outlier windows based on z-score from median.
Uses median and MAD (median absolute deviation) for robustness.
Args:
values: M values for each window
positions: (y, x, h, w) for each window
z_threshold: Z-score threshold for outlier detection
Returns:
Tuple of (low_outliers, high_outliers) - windows with unusually low/high M values
"""
arr = np.asarray(values, dtype=np.float64)
if arr.size < 5:
return [], []
median = float(np.median(arr))
mad = float(np.median(np.abs(arr - median)))
# Scale MAD to be consistent with std for normal distribution
mad_scaled = mad * 1.4826 if mad > 0 else 1e-10
low_outliers = []
high_outliers = []
for i, (val, pos) in enumerate(zip(values, positions)):
z_score = (val - median) / mad_scaled
if z_score < -z_threshold:
# Distinguish between zero M (flat region) and low M (weak CFA)
if val < 1e-6:
interp = "Zero CFA signal - likely flat/uniform region (sky, wall) or synthetic"
else:
interp = "Weak CFA signal - possible splice, synthetic region, or heavy processing"
low_outliers.append({
"y": pos[0],
"x": pos[1],
"h": pos[2],
"w": pos[3],
"M_value": float(val),
"z_score": float(z_score),
"interpretation": interp,
})
elif z_score > z_threshold:
high_outliers.append({
"y": pos[0],
"x": pos[1],
"h": pos[2],
"w": pos[3],
"M_value": float(val),
"z_score": float(z_score),
"interpretation": "Unusually strong CFA - possible different camera source",
})
# Sort by absolute z-score (most anomalous first)
low_outliers.sort(key=lambda x: x["z_score"])
high_outliers.sort(key=lambda x: -x["z_score"])
return low_outliers, high_outliers
def _classify_window_populations(
values: Sequence[float],
) -> Dict[str, Any]:
"""
Classify windows into populations based on M value magnitude.
Real camera images typically show:
- Low M (~0): Flat/uniform regions (sky, walls) - no texture to detect CFA
- High M (>1e9): Textured regions with strong CFA signal
This is content-dependent, not evidence of manipulation.
Manipulation would show as textured regions WITHOUT CFA signal.
"""
arr = np.asarray(values, dtype=np.float64)
if arr.size == 0:
return {"flat_regions": 0, "textured_regions": 0, "intermediate": 0}
# Thresholds based on typical M value ranges
# These are heuristic but based on observed values
flat_threshold = 1e6 # Below this = flat region (no texture)
textured_threshold = 1e9 # Above this = strong CFA signal
flat_count = int(np.sum(arr < flat_threshold))
textured_count = int(np.sum(arr >= textured_threshold))
intermediate_count = len(arr) - flat_count - textured_count
return {
"flat_regions": flat_count,
"textured_regions": textured_count,
"intermediate": intermediate_count,
"flat_pct": flat_count / len(arr) * 100,
"textured_pct": textured_count / len(arr) * 100,
}
def _window_brief(entry: Dict[str, Any], channel: str = "G") -> Dict[str, Any]:
"""Extract brief window info for a specific channel."""
return {
"y": entry["y"],
"x": entry["x"],
"h": entry["h"],
"w": entry["w"],
"M_value": float(entry[channel]["M"]),
}
def _analyze(params: Dict[str, Any]) -> Dict[str, Any]:
"""Run CFA consistency analysis on a single image."""
image_path = params.get("image_path")
if not image_path:
return {"error": "image_path is required for analyze mode."}
window = int(params.get("window", DEFAULT_WINDOW))
pattern = params.get("pattern", DEFAULT_PATTERN)
em_kwargs = params.get("em") or params.get("em_kwargs") or {}
top_k = int(params.get("top_k", DEFAULT_TOP_K))
channel = params.get("channel", "G").upper() # Green channel is most reliable
if channel not in ("R", "G", "B"):
channel = "G"
try:
img = cfa.load_rgb_image(str(image_path))
except Exception as e:
return {"error": f"Failed to load image: {e}"}
try:
window_results = cfa.analyze_image_windows(
img, window=window, pattern=pattern, em_kwargs=em_kwargs
)
except Exception as e:
return {"error": f"CFA analysis failed: {e}"}
if not window_results:
return {"error": "No windows analyzed (image may be too small)."}
# Extract M values and positions for the selected channel
m_values = [float(r[channel]["M"]) for r in window_results]
positions = [(r["y"], r["x"], r["h"], r["w"]) for r in window_results]
# Compute statistics
stats = _compute_stats(m_values)
# Analyze distribution
bimodality = _detect_bimodality(m_values)
# Classify windows into populations (flat vs textured)
populations = _classify_window_populations(m_values)
# Get top windows by M value (strongest CFA signal)
sorted_indices = np.argsort(m_values)[::-1]
top_windows = [
_window_brief(window_results[i], channel)
for i in sorted_indices[:top_k]
]
# Get bottom windows by M value (weakest CFA signal)
bottom_windows = [
_window_brief(window_results[i], channel)
for i in sorted_indices[-top_k:][::-1]
]
# Determine if image has CFA signal at all
has_cfa_signal = populations["textured_regions"] > 0
textured_pct = populations["textured_pct"]
# Generate interpretation based on content analysis
if not has_cfa_signal:
interpretation = (
"No strong CFA signal detected in any region. "
"This could indicate: (1) AI-generated image, (2) heavily processed image, "
"(3) screenshot, or (4) image with only flat/uniform content."
)
elif textured_pct > 50:
interpretation = (
f"Strong CFA signal detected in {textured_pct:.0f}% of windows. "
"Consistent with camera-captured image. Flat regions (sky, walls) "
"naturally show weaker CFA signal due to lack of texture."
)
elif textured_pct > 20:
interpretation = (
f"CFA signal detected in {textured_pct:.0f}% of windows (textured regions). "
"Remaining windows are flat/uniform regions where CFA cannot be detected. "
"This distribution is normal for photos with sky or uniform backgrounds."
)
else:
interpretation = (
f"Weak CFA signal - only {textured_pct:.0f}% of windows show strong CFA. "
"Image may be heavily processed, low-texture, or partially synthetic."
)
# Build result
result: Dict[str, Any] = {
"tool": "perform_cfa_detection",
"status": "completed",
"image_path": str(image_path),
"analysis_channel": channel,
"window_size": window,
"window_count": len(window_results),
"pattern": pattern,
# Main output: population analysis
"has_cfa_signal": has_cfa_signal,
"interpretation": interpretation,
# Window populations
"window_populations": {
"textured_with_cfa": populations["textured_regions"],
"flat_no_texture": populations["flat_regions"],
"intermediate": populations["intermediate"],
"textured_pct": populations["textured_pct"],
"flat_pct": populations["flat_pct"],
},
# Distribution analysis (for advanced users)
"distribution": {
"type": bimodality["distribution_type"],
"is_bimodal": bimodality["is_bimodal"],
"bimodality_coefficient": bimodality["bimodality_coefficient"],
"note": "Bimodal distribution is NORMAL for photos with mixed content (sky + texture)",
},
# Statistics
"m_value_stats": stats,
# Reference windows
"strongest_cfa_windows": top_windows,
"weakest_cfa_windows": bottom_windows,
"note": (
"CFA analysis detects demosaicing artifacts from camera sensors. "
"Flat regions (sky, walls) naturally have weak/no CFA signal. "
"Look for TEXTURED regions with weak CFA - those may be spliced. "
"This tool complements TruFor for localization, not whole-image classification."
),
}
return result
def _calibrate(params: Dict[str, Any]) -> Dict[str, Any]:
"""Calibrate reference statistics from a set of camera images."""
neg_dir = params.get("neg_dir") or params.get("ref_dir")
if not neg_dir:
return {"error": "neg_dir (or ref_dir) is required for calibrate mode."}
window = int(params.get("window", DEFAULT_WINDOW))
pattern = params.get("pattern", DEFAULT_PATTERN)
em_kwargs = params.get("em") or params.get("em_kwargs") or {}
save_to = params.get("save_to") or params.get("output")
neg_files = cfa.list_image_files(str(neg_dir))
if not neg_files:
return {"error": f"No images found in directory: {neg_dir}"}
# Collect M values from all reference images
all_m_values: Dict[str, List[float]] = {"R": [], "G": [], "B": []}
for path in neg_files:
try:
img = cfa.load_rgb_image(str(path))
window_results = cfa.analyze_image_windows(
img, window=window, pattern=pattern, em_kwargs=em_kwargs
)
for r in window_results:
all_m_values["R"].append(float(r["R"]["M"]))
all_m_values["G"].append(float(r["G"]["M"]))
all_m_values["B"].append(float(r["B"]["M"]))
except Exception:
continue # Skip problematic images
if not all_m_values["G"]:
return {"error": "No valid windows collected from reference images."}
# Compute reference statistics for each channel
reference_stats = {
c: _compute_stats(vals) for c, vals in all_m_values.items()
}
payload = {
"reference_stats": reference_stats,
"pattern": pattern,
"window": window,
"em_params": em_kwargs,
"num_images": len(neg_files),
"num_windows": len(all_m_values["G"]),
}
if save_to:
Path(save_to).write_text(json.dumps(payload, indent=2), encoding="utf-8")
payload["saved_to"] = str(save_to)
return payload
def perform_cfa_detection(input_str: str) -> str:
"""
LangChain tool entrypoint for CFA consistency analysis.
This tool analyzes CFA (Color Filter Array) demosaicing patterns to detect
INCONSISTENCIES within an image. It is designed for splice detection and
source consistency analysis.
Input (JSON):
- mode: "analyze" (default) or "calibrate"
- image_path: required for analyze
- window: int (default 256)
- pattern: Bayer pattern (default RGGB)
- channel: which channel to analyze (default "G" - green is most reliable)
- em / em_kwargs: dict for EM params (N, sigma0, p0, max_iter, tol, seed)
- top_k: int (default 5) - number of top/outlier windows to return
- outlier_zscore: float (default 2.0) - z-score threshold for outlier detection
- neg_dir/ref_dir: required for calibrate mode
- save_to: optional path to write reference stats JSON (calibrate)
Output:
- cfa_consistency_score: 0-1 score (higher = more consistent)
- distribution: analysis of M value distribution (unimodal/bimodal)
- outliers: windows with unusually low/high CFA patterns
- interpretation: human-readable summary
"""
params = _parse_request(input_str)
mode = params.get("mode", "analyze").lower()
# Support legacy "detect" mode name
if mode == "detect":
mode = "analyze"
if mode == "calibrate":
result = _calibrate(params)
elif mode == "analyze":
result = _analyze(params)
else:
result = {"error": "mode must be 'analyze' or 'calibrate'."}
try:
return json.dumps(result, indent=2)
except Exception:
# Fallback in case something is not JSON-serializable
return json.dumps({"error": "Failed to serialize result."}, indent=2)
__all__ = ["perform_cfa_detection"]