Datasets:

Vereye
/

benchmark-synthid-image-protocol

#!/usr/bin/env python3
"""run_competitors.py - re-runs the WAM and TrustMark rows of README.md on this folder's photos and attacks, and compares every value
with sets/<set>/results/rivals.csv.
Each method marks each set's ORIGINALS (rebuilt from the public datasets by scripts/prep.py: run scripts/download_sources.sh first) with
its own message (bits_for below: the benchmark's message for that photo), merging its watermark into the full-size photo BICUBIC (the
SynthID-Image paper's choice); the folder's transforms.py attacks the marked copy and the unmarked original with the same seeds as the
Vereye copies; every copy - marked AND unmarked - is decoded against the photo's own message. Scores: WAM's bit accuracy and its own
detection (mask) score; TrustMark's data-bit accuracy after its error correction, its raw 100-bit accuracy before it, and its presence flag.
Then: python3 scripts/score.py on a copy of the folder whose rivals.csv is this script's output gives the README's rival rows.
The exact versions used: torch 2.14.1 (CPU), torchvision 0.29.1, trustmark 0.9.2 (its Q models: decoder_Q.ckpt SHA-256
e3d9cea5406a26590735719f8f15cb10802b11852ae69047eaf4cf17214df781, encoder_Q.ckpt dc382c3f6b4fd568b27d6fbb763d6ffc2d2587126d84afe9d5ee95b4c5d99826),
Watermark Anything at commit 2c08af04d037d5667c02f6ddebbda9ff04581c3e with checkpoints/wam_mit.pth SHA-256
90ef232384e023bd63245eb0c131abd69d2afc7b8f17a71ccedceb542bf009e2. Our published rival rows came from a batched run of the same arithmetic
(many copies per network pass; checked equal to this per-image decode on 12 copies, marked and unmarked, WAM and TrustMark).
  pip install -r requirements.txt torch==2.14.1 torchvision==0.29.1 trustmark==0.9.2 omegaconf
  git clone https://github.com/facebookresearch/watermark-anything && git -C watermark-anything checkout 2c08af04d037d5667c02f6ddebbda9ff04581c3e
  python3 scripts/run_competitors.py --sources DIR --set heldout-coco-val2017-512 --method trustmark --photos 000000570448,000000570456
  python3 scripts/run_competitors.py --sources DIR --set heldout-coco-val2017-512 --method wam --wam-code watermark-anything
The output (rerun-<method>-<set>.csv) is written to the CURRENT directory, never into this folder."""
import argparse, csv, hashlib, io, json, os, sys
import numpy as np
from PIL import Image
HERE = os.path.dirname(os.path.abspath(__file__)); KIT = os.path.dirname(HERE); sys.path.insert(0, HERE)
import transforms as K, prep  # noqa: E402

def bits_for(img_id, n, salt):
  h = hashlib.sha256(f'{salt}:{img_id}'.encode()).digest() * 4
  return [(h[i // 8] >> (i % 8)) & 1 for i in range(n)]

class TrustMarkQ:
  salt = 'tm'
  def __init__(self, _):
    import torch; torch.set_num_threads(4)
    from trustmark import TrustMark
    self.tm = TrustMark(verbose=False, model_type='Q', encoding_type=TrustMark.Encoding.BCH_5); self.n = self.tm.schemaCapacity()
  def embed(self, im, bits):
    s = ''.join(map(str, bits)); self.cw = np.asarray(self.tm.ecc.encode_binary([s])).ravel().astype(int)
    return self.tm.encode(im, s, MODE='binary', WM_MERGE='bicubic')
  def decode(self, im, bits):
    import torch
    from torchvision import transforms as TV
    secret, present, _ = self.tm.decode(im, MODE='binary'); got = [int(c) for c in secret][:self.n] if secret else []
    r = self.tm.model_resolution_dec; x = TV.ToTensor()(self.tm.get_the_image_for_processing(im).resize((r, r), Image.BILINEAR)).unsqueeze(0) * 2.0 - 1.0
    with torch.no_grad(): raw = (self.tm.decoder.decoder(x) > 0).cpu().numpy().ravel().astype(int)
    return {'bitacc': acc(got, bits), 'rawacc': float(np.mean(raw[:len(self.cw)] == self.cw)), 'mask': '', 'present': bool(present), 'ident': bool(present and got == bits)}

class WAM:
  salt = 'wam'
  def __init__(self, code):
    import torch, omegaconf; torch.set_num_threads(4); sys.path.insert(0, code); cwd = os.getcwd(); os.chdir(code)
    try:
      from watermark_anything.models import Wam, build_embedder, build_extractor
      from watermark_anything.augmentation.augmenter import Augmenter
      from watermark_anything.data.transforms import default_transform, normalize_img, unnormalize_img
      from watermark_anything.modules.jnd import JND
      from watermark_anything.data.metrics import msg_predict_inference
      a = argparse.Namespace(**json.load(open('checkpoints/params.json')))
      emb = omegaconf.OmegaConf.load(a.embedder_config)[a.embedder_model]; ext = omegaconf.OmegaConf.load(a.extractor_config)
      att = JND(**omegaconf.OmegaConf.load(a.attenuation_config)[a.attenuation], preprocess=unnormalize_img, postprocess=normalize_img)
      self.wam = Wam(build_embedder(a.embedder_model, emb, a.nbits), build_extractor(ext.model, ext[a.extractor_model], a.img_size, a.nbits),
                     Augmenter(**omegaconf.OmegaConf.load(a.augmentation_config)), att, a.scaling_w, a.scaling_i)
      self.wam.load_state_dict(torch.load('checkpoints/wam_mit.pth', map_location='cpu')); self.wam.eval()
    finally: os.chdir(cwd)
    self.torch, self.T, self.unnorm, self.predict, self.n = torch, default_transform, unnormalize_img, msg_predict_inference, a.nbits
  def embed(self, im, bits):
    from torchvision import transforms as TV
    with self.torch.no_grad():
      x = self.T(im).unsqueeze(0); h = self.wam.img_size_extractor
      r = TV.Resize((h, h), interpolation=TV.InterpolationMode.BILINEAR)(x) if x.shape[-2:] != (h, h) else x
      pw = self.wam.embedder(r, self.torch.tensor(bits, dtype=self.torch.float32).unsqueeze(0))
      if x.shape[-2:] != (h, h): pw = TV.Resize(x.shape[-2:], interpolation=TV.InterpolationMode.BICUBIC, antialias=True)(pw)
      return Image.fromarray(self.unnorm(self.wam.blend(x, pw)[0]).clamp(0, 1).mul(255).round().byte().permute(1, 2, 0).numpy())
  def decode(self, im, bits):
    F = self.torch.nn.functional
    with self.torch.no_grad():
      p = self.wam.detect(self.T(im).unsqueeze(0))['preds']; m = F.sigmoid(p[:, 0:1]); got = [int(x) for x in self.predict(p[:, 1:], m).float()[0].tolist()]
    return {'bitacc': acc(got, bits), 'rawacc': '', 'mask': float(m.mean()), 'present': '', 'ident': got == bits}

def png(im): b = io.BytesIO(); im.save(b, 'PNG', compress_level=1); b.seek(0); return Image.open(b).convert('RGB')  # lossless, as the benchmark kept them
def acc(got, ref): return sum(int(a == b) for a, b in zip(got, ref)) / len(ref) if len(got) == len(ref) else 0.0

if __name__ == '__main__':
  ap = argparse.ArgumentParser(); ap.add_argument('--sources', required=True); ap.add_argument('--set', required=True)
  ap.add_argument('--method', choices=['wam', 'trustmark'], required=True); ap.add_argument('--wam-code', default='watermark-anything'); ap.add_argument('--photos'); a = ap.parse_args()
  D = os.path.join(KIT, 'sets', a.set); K.S = json.load(open(os.path.join(D, 'set.json')))['size']; M = (WAM if a.method == 'wam' else TrustMarkQ)(a.wam_code)
  photos = sorted(os.listdir(os.path.join(D, 'photos')))
  if a.photos: photos = [p for p in photos if p in set(a.photos.split(','))]
  pub = {(r['photo'], r['transformation'], r['mode'], r['marked']): r for r in csv.DictReader(open(os.path.join(D, 'results', 'rivals.csv'), newline='')) if r['method'] == a.method}
  out = os.path.abspath(f'rerun-{a.method}-{a.set}.csv'); rows = []; same = diff = 0; TMP = __import__('tempfile').mkdtemp()
  for p in photos:
    info = json.load(open(os.path.join(D, 'photos', p, 'info.json'))); orig = prep.prep(os.path.join(a.sources, info['source_file']['file']), os.path.join(TMP, f'{p}.png'), K.S)
    os.remove(os.path.join(TMP, f'{p}.png')); bits = bits_for(p, M.n, M.salt); marked = png(M.embed(orig, bits))
    for n in K.T:
      for md in ('random', 'worst'):
        for mk, src in ((1, marked), (0, orig)):
          im, _ = K.apply(n, md, src, K.seed_of(p, n, md)); r = {'set': a.set, 'method': a.method, 'photo': p, 'transformation': n, 'mode': md, 'group': K.T[n][0], 'marked': mk, **M.decode(png(im), bits)}
          rows.append(r); q = pub.get((p, n, md, str(mk)))
          if q: ok = abs(float(q['bitacc'] or 0) - float(r['bitacc'] or 0)) < 1e-9; same += ok; diff += not ok
    print(f'{p} done', flush=True)
  with open(out, 'w', newline='') as f: w = csv.DictWriter(f, list(rows[0])); w.writeheader(); w.writerows(rows)
  print(f'{len(rows)} copies scored -> {out}; bit accuracy equal to results/rivals.csv on {same} of {same + diff} copies')

	#!/usr/bin/env python3
	"""run_competitors.py - re-runs the WAM and TrustMark rows of README.md on this folder's photos and attacks, and compares every value
	with sets/<set>/results/rivals.csv.

	Each method marks each set's ORIGINALS (rebuilt from the public datasets by scripts/prep.py: run scripts/download_sources.sh first) with
	its own message (bits_for below: the benchmark's message for that photo), merging its watermark into the full-size photo BICUBIC (the
	SynthID-Image paper's choice); the folder's transforms.py attacks the marked copy and the unmarked original with the same seeds as the
	Vereye copies; every copy - marked AND unmarked - is decoded against the photo's own message. Scores: WAM's bit accuracy and its own
	detection (mask) score; TrustMark's data-bit accuracy after its error correction, its raw 100-bit accuracy before it, and its presence flag.
	Then: python3 scripts/score.py on a copy of the folder whose rivals.csv is this script's output gives the README's rival rows.

	The exact versions used: torch 2.14.1 (CPU), torchvision 0.29.1, trustmark 0.9.2 (its Q models: decoder_Q.ckpt SHA-256
	e3d9cea5406a26590735719f8f15cb10802b11852ae69047eaf4cf17214df781, encoder_Q.ckpt dc382c3f6b4fd568b27d6fbb763d6ffc2d2587126d84afe9d5ee95b4c5d99826),
	Watermark Anything at commit 2c08af04d037d5667c02f6ddebbda9ff04581c3e with checkpoints/wam_mit.pth SHA-256
	90ef232384e023bd63245eb0c131abd69d2afc7b8f17a71ccedceb542bf009e2. Our published rival rows came from a batched run of the same arithmetic
	(many copies per network pass; checked equal to this per-image decode on 12 copies, marked and unmarked, WAM and TrustMark).
	pip install -r requirements.txt torch==2.14.1 torchvision==0.29.1 trustmark==0.9.2 omegaconf
	git clone https://github.com/facebookresearch/watermark-anything && git -C watermark-anything checkout 2c08af04d037d5667c02f6ddebbda9ff04581c3e
	python3 scripts/run_competitors.py --sources DIR --set heldout-coco-val2017-512 --method trustmark --photos 000000570448,000000570456
	python3 scripts/run_competitors.py --sources DIR --set heldout-coco-val2017-512 --method wam --wam-code watermark-anything
	The output (rerun-<method>-<set>.csv) is written to the CURRENT directory, never into this folder."""
	import argparse, csv, hashlib, io, json, os, sys
	import numpy as np
	from PIL import Image
	HERE = os.path.dirname(os.path.abspath(__file__)); KIT = os.path.dirname(HERE); sys.path.insert(0, HERE)
	import transforms as K, prep # noqa: E402

	def bits_for(img_id, n, salt):
	h = hashlib.sha256(f'{salt}:{img_id}'.encode()).digest() * 4
	return [(h[i // 8] >> (i % 8)) & 1 for i in range(n)]

	class TrustMarkQ:
	salt = 'tm'
	def __init__(self, _):
	import torch; torch.set_num_threads(4)
	from trustmark import TrustMark
	self.tm = TrustMark(verbose=False, model_type='Q', encoding_type=TrustMark.Encoding.BCH_5); self.n = self.tm.schemaCapacity()
	def embed(self, im, bits):
	s = ''.join(map(str, bits)); self.cw = np.asarray(self.tm.ecc.encode_binary([s])).ravel().astype(int)
	return self.tm.encode(im, s, MODE='binary', WM_MERGE='bicubic')
	def decode(self, im, bits):
	import torch
	from torchvision import transforms as TV
	secret, present, _ = self.tm.decode(im, MODE='binary'); got = [int(c) for c in secret][:self.n] if secret else []
	r = self.tm.model_resolution_dec; x = TV.ToTensor()(self.tm.get_the_image_for_processing(im).resize((r, r), Image.BILINEAR)).unsqueeze(0) * 2.0 - 1.0
	with torch.no_grad(): raw = (self.tm.decoder.decoder(x) > 0).cpu().numpy().ravel().astype(int)
	return {'bitacc': acc(got, bits), 'rawacc': float(np.mean(raw[:len(self.cw)] == self.cw)), 'mask': '', 'present': bool(present), 'ident': bool(present and got == bits)}

	class WAM:
	salt = 'wam'
	def __init__(self, code):
	import torch, omegaconf; torch.set_num_threads(4); sys.path.insert(0, code); cwd = os.getcwd(); os.chdir(code)
	try:
	from watermark_anything.models import Wam, build_embedder, build_extractor
	from watermark_anything.augmentation.augmenter import Augmenter
	from watermark_anything.data.transforms import default_transform, normalize_img, unnormalize_img
	from watermark_anything.modules.jnd import JND
	from watermark_anything.data.metrics import msg_predict_inference
	a = argparse.Namespace(**json.load(open('checkpoints/params.json')))
	emb = omegaconf.OmegaConf.load(a.embedder_config)[a.embedder_model]; ext = omegaconf.OmegaConf.load(a.extractor_config)
	att = JND(**omegaconf.OmegaConf.load(a.attenuation_config)[a.attenuation], preprocess=unnormalize_img, postprocess=normalize_img)
	self.wam = Wam(build_embedder(a.embedder_model, emb, a.nbits), build_extractor(ext.model, ext[a.extractor_model], a.img_size, a.nbits),
	Augmenter(**omegaconf.OmegaConf.load(a.augmentation_config)), att, a.scaling_w, a.scaling_i)
	self.wam.load_state_dict(torch.load('checkpoints/wam_mit.pth', map_location='cpu')); self.wam.eval()
	finally: os.chdir(cwd)
	self.torch, self.T, self.unnorm, self.predict, self.n = torch, default_transform, unnormalize_img, msg_predict_inference, a.nbits
	def embed(self, im, bits):
	from torchvision import transforms as TV
	with self.torch.no_grad():
	x = self.T(im).unsqueeze(0); h = self.wam.img_size_extractor
	r = TV.Resize((h, h), interpolation=TV.InterpolationMode.BILINEAR)(x) if x.shape[-2:] != (h, h) else x
	pw = self.wam.embedder(r, self.torch.tensor(bits, dtype=self.torch.float32).unsqueeze(0))
	if x.shape[-2:] != (h, h): pw = TV.Resize(x.shape[-2:], interpolation=TV.InterpolationMode.BICUBIC, antialias=True)(pw)
	return Image.fromarray(self.unnorm(self.wam.blend(x, pw)[0]).clamp(0, 1).mul(255).round().byte().permute(1, 2, 0).numpy())
	def decode(self, im, bits):
	F = self.torch.nn.functional
	with self.torch.no_grad():
	p = self.wam.detect(self.T(im).unsqueeze(0))['preds']; m = F.sigmoid(p[:, 0:1]); got = [int(x) for x in self.predict(p[:, 1:], m).float()[0].tolist()]
	return {'bitacc': acc(got, bits), 'rawacc': '', 'mask': float(m.mean()), 'present': '', 'ident': got == bits}

	def png(im): b = io.BytesIO(); im.save(b, 'PNG', compress_level=1); b.seek(0); return Image.open(b).convert('RGB') # lossless, as the benchmark kept them
	def acc(got, ref): return sum(int(a == b) for a, b in zip(got, ref)) / len(ref) if len(got) == len(ref) else 0.0

	if __name__ == '__main__':
	ap = argparse.ArgumentParser(); ap.add_argument('--sources', required=True); ap.add_argument('--set', required=True)
	ap.add_argument('--method', choices=['wam', 'trustmark'], required=True); ap.add_argument('--wam-code', default='watermark-anything'); ap.add_argument('--photos'); a = ap.parse_args()
	D = os.path.join(KIT, 'sets', a.set); K.S = json.load(open(os.path.join(D, 'set.json')))['size']; M = (WAM if a.method == 'wam' else TrustMarkQ)(a.wam_code)
	photos = sorted(os.listdir(os.path.join(D, 'photos')))
	if a.photos: photos = [p for p in photos if p in set(a.photos.split(','))]
	pub = {(r['photo'], r['transformation'], r['mode'], r['marked']): r for r in csv.DictReader(open(os.path.join(D, 'results', 'rivals.csv'), newline='')) if r['method'] == a.method}
	out = os.path.abspath(f'rerun-{a.method}-{a.set}.csv'); rows = []; same = diff = 0; TMP = __import__('tempfile').mkdtemp()
	for p in photos:
	info = json.load(open(os.path.join(D, 'photos', p, 'info.json'))); orig = prep.prep(os.path.join(a.sources, info['source_file']['file']), os.path.join(TMP, f'{p}.png'), K.S)
	os.remove(os.path.join(TMP, f'{p}.png')); bits = bits_for(p, M.n, M.salt); marked = png(M.embed(orig, bits))
	for n in K.T:
	for md in ('random', 'worst'):
	for mk, src in ((1, marked), (0, orig)):
	im, _ = K.apply(n, md, src, K.seed_of(p, n, md)); r = {'set': a.set, 'method': a.method, 'photo': p, 'transformation': n, 'mode': md, 'group': K.T[n][0], 'marked': mk, **M.decode(png(im), bits)}
	rows.append(r); q = pub.get((p, n, md, str(mk)))
	if q: ok = abs(float(q['bitacc'] or 0) - float(r['bitacc'] or 0)) < 1e-9; same += ok; diff += not ok
	print(f'{p} done', flush=True)
	with open(out, 'w', newline='') as f: w = csv.DictWriter(f, list(rows[0])); w.writeheader(); w.writerows(rows)
	print(f'{len(rows)} copies scored -> {out}; bit accuracy equal to results/rivals.csv on {same} of {same + diff} copies')