File size: 40,081 Bytes
61f8287 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 | {
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 🔍 AI-Generated Image Detector — Kaggle Training\n",
"\n",
"**Architecture:** SwinV2 + SRM High-Pass Filters + DCT + FFT Frequency Analysis\n",
"\n",
"**Dataset:** tristanzhang32/ai-generated-images-vs-real-images (60K images)\n",
"\n",
"**Hardware:** T4 x2 (2× 16GB VRAM)\n",
"\n",
"---\n",
"\n",
"### Setup Instructions\n",
"1. **Add Dataset:** `tristanzhang32/ai-generated-images-vs-real-images`\n",
"2. **Accelerator:** GPU T4 x2\n",
"3. **Internet:** ON (needed for SwinV2 weights + pushing to HF Hub)\n",
"4. **Kaggle Secret:** Add `HF_TOKEN` with your Hugging Face write token (for push_to_hub)\n",
"5. **Run All**"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"_cell_guid": "b1076dfc-b9ad-4769-8c92-a6c4dae69d19",
"_uuid": "8f2839f25d086af736a60e9eeb907d3b93b6e0e5",
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 1: Install dependencies\n",
"# ============================================================\n",
"!pip install -q transformers accelerate evaluate trackio scikit-learn huggingface_hub"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 2: Imports + HF Login\n",
"# ============================================================\n",
"import os, io, math, random, json, glob\n",
"import numpy as np\n",
"from pathlib import Path\n",
"from PIL import Image, ImageFilter\n",
"\n",
"import torch\n",
"import torch.nn as nn\n",
"import torch.nn.functional as F\n",
"from torch.utils.data import Dataset, DataLoader, random_split\n",
"\n",
"from transformers import (\n",
" Swinv2Model,\n",
" TrainingArguments,\n",
" Trainer,\n",
")\n",
"from torchvision.transforms import (\n",
" Compose, Normalize, Resize, CenterCrop, RandomResizedCrop,\n",
" RandomHorizontalFlip, ToTensor, ColorJitter,\n",
")\n",
"import evaluate\n",
"import trackio\n",
"\n",
"# Login to HF Hub (uses Kaggle Secret)\n",
"from huggingface_hub import login\n",
"from kaggle_secrets import UserSecretsClient\n",
"try:\n",
" secrets = UserSecretsClient()\n",
" hf_token = secrets.get_secret(\"HF_TOKEN\")\n",
" login(token=hf_token)\n",
" print(\"✓ Logged in to Hugging Face Hub\")\n",
"except Exception as e:\n",
" print(f\"⚠ Could not login to HF Hub: {e}\")\n",
" print(\" Model will be saved locally only. Add HF_TOKEN to Kaggle Secrets to push to Hub.\")\n",
"\n",
"print(f\"PyTorch: {torch.__version__}\")\n",
"print(f\"CUDA available: {torch.cuda.is_available()}\")\n",
"print(f\"GPUs: {torch.cuda.device_count()}\")\n",
"for i in range(torch.cuda.device_count()):\n",
" print(f\" GPU {i}: {torch.cuda.get_device_name(i)} ({torch.cuda.get_device_properties(i).total_mem / 1e9:.1f} GB)\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 3: Explore dataset structure\n",
"# ============================================================\n",
"BASE_PATH = Path(\"/kaggle/input/ai-generated-images-vs-real-images\")\n",
"\n",
"# Auto-detect folder structure\n",
"if not BASE_PATH.exists():\n",
" # Try alternative paths\n",
" candidates = list(Path(\"/kaggle/input\").glob(\"**/real\")) + list(Path(\"/kaggle/input\").glob(\"**/Real\"))\n",
" if candidates:\n",
" BASE_PATH = candidates[0].parent\n",
" print(f\"Found dataset at: {BASE_PATH}\")\n",
" else:\n",
" # List everything in /kaggle/input\n",
" for p in sorted(Path(\"/kaggle/input\").rglob(\"*\")):\n",
" if p.is_dir():\n",
" n_files = len(list(p.iterdir()))\n",
" print(f\" DIR {p} ({n_files} items)\")\n",
" raise FileNotFoundError(\"Could not find dataset. Check the paths above and update BASE_PATH.\")\n",
"\n",
"print(f\"\\nDataset root: {BASE_PATH}\")\n",
"print(f\"Contents:\")\n",
"for item in sorted(BASE_PATH.iterdir()):\n",
" if item.is_dir():\n",
" n_files = len(list(item.rglob(\"*.*\")))\n",
" print(f\" 📁 {item.name}/ ({n_files} files)\")\n",
" else:\n",
" print(f\" 📄 {item.name} ({item.stat().st_size / 1e6:.1f} MB)\")\n",
"\n",
"# Detect class folders\n",
"subdirs = sorted([d.name for d in BASE_PATH.iterdir() if d.is_dir()])\n",
"print(f\"\\nSubdirectories: {subdirs}\")\n",
"\n",
"# Check for nested structure (e.g., train/real, train/fake)\n",
"for sd in subdirs:\n",
" sub_subdirs = sorted([d.name for d in (BASE_PATH / sd).iterdir() if d.is_dir()])\n",
" if sub_subdirs:\n",
" print(f\" {sd}/ contains: {sub_subdirs}\")\n",
" for ssd in sub_subdirs:\n",
" n = len(list((BASE_PATH / sd / ssd).rglob(\"*.*\")))\n",
" print(f\" {ssd}/: {n} files\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 4: SRM Filter Bank (30 forensic high-pass filters)\n",
"# ============================================================\n",
"\n",
"def get_srm_kernels():\n",
" \"\"\"30 SRM high-pass filter kernels (5x5) for forensic residual extraction.\n",
" Based on Fridrich & Kodovsky (2012) - Rich Models for Steganalysis.\n",
" These detect manipulation artifacts invisible in RGB space.\"\"\"\n",
" f1 = np.array([[0,0,0,0,0],[0,0,0,0,0],[0,0,-1,1,0],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" f2 = np.array([[0,0,0,0,0],[0,0,0,0,0],[0,0,-1,0,0],[0,0,1,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" f3 = np.array([[0,0,0,0,0],[0,0,0,0,0],[0,1,-2,1,0],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" f4 = np.array([[0,0,0,0,0],[0,0,1,0,0],[0,0,-2,0,0],[0,0,1,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" f5 = np.array([[0,0,0,0,0],[0,1,0,0,0],[0,0,-2,0,0],[0,0,0,1,0],[0,0,0,0,0]], dtype=np.float32)\n",
" f6 = np.array([[0,0,0,0,0],[0,0,0,1,0],[0,0,-2,0,0],[0,1,0,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" f7 = np.array([[0,0,0,0,0],[0,0,0,0,0],[0,-1,3,-3,1],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" f8 = np.array([[0,0,0,0,0],[0,0,-1,0,0],[0,0,3,0,0],[0,0,-3,0,0],[0,0,1,0,0]], dtype=np.float32)\n",
" f9 = np.array([[0,0,0,0,0],[0,0,1,0,0],[0,1,-4,1,0],[0,0,1,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" f10 = np.array([[0,0,0,0,0],[0,1,1,1,0],[0,1,-8,1,0],[0,1,1,1,0],[0,0,0,0,0]], dtype=np.float32) / 3.0\n",
" f11 = np.array([[0,0,0,0,0],[0,-1,2,-1,0],[0,2,-4,2,0],[0,-1,2,-1,0],[0,0,0,0,0]], dtype=np.float32)\n",
" f12 = np.array([[0,0,0,0,0],[0,0,0,0,0],[-1,2,-2,2,-1],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32) / 2.0\n",
" f13 = np.array([[0,0,-1,0,0],[0,0,2,0,0],[0,0,-2,0,0],[0,0,2,0,0],[0,0,-1,0,0]], dtype=np.float32) / 2.0\n",
" f14 = np.array([[0,0,-1,0,0],[0,0,2,0,0],[-1,2,-4,2,-1],[0,0,2,0,0],[0,0,-1,0,0]], dtype=np.float32) / 4.0\n",
" f15 = np.array([[-1,2,-2,2,-1],[2,-6,8,-6,2],[-2,8,-12,8,-2],[2,-6,8,-6,2],[-1,2,-2,2,-1]], dtype=np.float32) / 12.0\n",
" spam_h = np.array([[0,0,0,0,0],[0,0,0,0,0],[0,-1,2,-1,0],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" spam_v = np.array([[0,0,0,0,0],[0,0,-1,0,0],[0,0,2,0,0],[0,0,-1,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" spam_d1 = np.array([[0,0,0,0,0],[0,-1,0,0,0],[0,0,2,0,0],[0,0,0,-1,0],[0,0,0,0,0]], dtype=np.float32)\n",
" spam_d2 = np.array([[0,0,0,0,0],[0,0,0,-1,0],[0,0,2,0,0],[0,-1,0,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" spam3_h = np.array([[0,0,0,0,0],[0,0,0,0,0],[1,-3,3,-1,0],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" spam3_v = np.array([[0,0,1,0,0],[0,0,-3,0,0],[0,0,3,0,0],[0,0,-1,0,0],[0,0,0,0,0]], dtype=np.float32)\n",
" sq5_1 = np.array([[0,0,0,0,0],[0,0,1,0,0],[0,1,-4,1,0],[0,0,1,0,0],[0,0,0,0,0]], dtype=np.float32) / 2.0\n",
" sq5_2 = np.array([[0,0,0,0,0],[0,1,0,1,0],[0,0,-4,0,0],[0,1,0,1,0],[0,0,0,0,0]], dtype=np.float32) / 2.0\n",
" cross1 = np.array([[0,0,-1,0,0],[0,0,2,0,0],[0,0,-2,0,0],[0,0,2,0,0],[0,0,-1,0,0]], dtype=np.float32) / 2.0\n",
" cross2 = np.array([[0,0,0,0,0],[0,0,0,0,0],[-1,2,-2,2,-1],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32) / 2.0\n",
" edge_d1 = np.array([[-1,0,0,0,0],[0,2,0,0,0],[0,0,-2,0,0],[0,0,0,2,0],[0,0,0,0,-1]], dtype=np.float32) / 2.0\n",
" edge_d2 = np.array([[0,0,0,0,-1],[0,0,0,2,0],[0,0,-2,0,0],[0,2,0,0,0],[-1,0,0,0,0]], dtype=np.float32) / 2.0\n",
" gabor_h = np.array([[0,0,0,0,0],[1,-1,0,-1,1],[0,0,0,0,0],[-1,1,0,1,-1],[0,0,0,0,0]], dtype=np.float32) / 4.0\n",
" gabor_v = np.array([[0,1,0,-1,0],[0,-1,0,1,0],[0,0,0,0,0],[0,-1,0,1,0],[0,1,0,-1,0]], dtype=np.float32) / 4.0\n",
" all_filters = [f1,f2,f3,f4,f5,f6,f7,f8,f9,f10,f11,f12,f13,f14,f15,\n",
" spam_h,spam_v,spam_d1,spam_d2,spam3_h,spam3_v,\n",
" sq5_1,sq5_2,cross1,cross2,edge_d1,edge_d2,gabor_h,gabor_v,f15]\n",
" return all_filters[:30]\n",
"\n",
"\n",
"class SRMFilterBank(nn.Module):\n",
" \"\"\"30 fixed forensic high-pass filters. No learnable parameters.\"\"\"\n",
" def __init__(self):\n",
" super().__init__()\n",
" kernels = get_srm_kernels()\n",
" weight = torch.stack([torch.tensor(k) for k in kernels]).unsqueeze(1)\n",
" weight = weight.repeat(1, 3, 1, 1) / 3.0\n",
" self.register_buffer('weight', weight)\n",
" def forward(self, x):\n",
" return F.conv2d(x, self.weight, padding=2)\n",
"\n",
"print(\"✓ SRM Filter Bank defined (30 kernels)\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 5: DCT Frequency Analyzer\n",
"# ============================================================\n",
"\n",
"class DCTFrequencyAnalyzer(nn.Module):\n",
" \"\"\"2D DCT on image patches → frequency band energy statistics.\n",
" AI images show different frequency distributions than natural images.\"\"\"\n",
" def __init__(self, patch_size=32, num_freq_bands=8):\n",
" super().__init__()\n",
" self.patch_size = patch_size\n",
" self.num_freq_bands = num_freq_bands\n",
" N = patch_size\n",
" dct_mat = torch.zeros(N, N)\n",
" for k in range(N):\n",
" for n in range(N):\n",
" if k == 0:\n",
" dct_mat[k, n] = math.sqrt(1.0 / N)\n",
" else:\n",
" dct_mat[k, n] = math.sqrt(2.0 / N) * math.cos(math.pi * (2*n + 1) * k / (2*N))\n",
" self.register_buffer('dct_mat', dct_mat)\n",
" self.register_buffer('dct_mat_t', dct_mat.t())\n",
"\n",
" def dct2d(self, x):\n",
" return torch.matmul(torch.matmul(self.dct_mat, x), self.dct_mat_t)\n",
"\n",
" def forward(self, x):\n",
" B, C, H, W = x.shape\n",
" ps = self.patch_size\n",
" gray = 0.299 * x[:, 0] + 0.587 * x[:, 1] + 0.114 * x[:, 2]\n",
" h_patches = H // ps\n",
" w_patches = W // ps\n",
" gray = gray[:, :h_patches*ps, :w_patches*ps]\n",
" patches = gray.unfold(1, ps, ps).unfold(2, ps, ps)\n",
" B_p, hp, wp = patches.shape[:3]\n",
" patches = patches.reshape(B_p * hp * wp, ps, ps)\n",
" dct_patches = self.dct2d(patches)\n",
" dct_patches = dct_patches.reshape(B_p, hp * wp, ps, ps)\n",
"\n",
" features = []\n",
" freq_y = torch.arange(ps, device=x.device).float()\n",
" freq_x = torch.arange(ps, device=x.device).float()\n",
" fy, fx = torch.meshgrid(freq_y, freq_x, indexing='ij')\n",
" freq_dist = torch.sqrt(fy**2 + fx**2)\n",
" max_freq = math.sqrt(2) * ps\n",
"\n",
" for band in range(self.num_freq_bands):\n",
" lo = band * max_freq / self.num_freq_bands\n",
" hi = (band + 1) * max_freq / self.num_freq_bands\n",
" mask = ((freq_dist >= lo) & (freq_dist < hi)).float()\n",
" band_energy = (dct_patches ** 2 * mask.unsqueeze(0).unsqueeze(0)).sum(dim=(-2, -1))\n",
" features.append(band_energy.mean(dim=1, keepdim=True))\n",
" features.append(band_energy.std(dim=1, keepdim=True))\n",
"\n",
" total_energy = (dct_patches ** 2).sum(dim=(-2, -1))\n",
" weighted_freq = (dct_patches ** 2 * freq_dist.unsqueeze(0).unsqueeze(0)).sum(dim=(-2, -1))\n",
" spectral_centroid = weighted_freq / (total_energy + 1e-8)\n",
" features.append(spectral_centroid.mean(dim=1, keepdim=True))\n",
" features.append(spectral_centroid.std(dim=1, keepdim=True))\n",
"\n",
" mid = ps // 2\n",
" low_mask = (freq_dist < mid).float()\n",
" high_mask = (freq_dist >= mid).float()\n",
" low_energy = (dct_patches ** 2 * low_mask).sum(dim=(-2, -1))\n",
" high_energy = (dct_patches ** 2 * high_mask).sum(dim=(-2, -1))\n",
" hl_ratio = high_energy / (low_energy + 1e-8)\n",
" features.append(hl_ratio.mean(dim=1, keepdim=True))\n",
" features.append(hl_ratio.std(dim=1, keepdim=True))\n",
"\n",
" dc_values = dct_patches[:, :, 0, 0]\n",
" features.append(dc_values.mean(dim=1, keepdim=True))\n",
" features.append(dc_values.std(dim=1, keepdim=True))\n",
"\n",
" return torch.cat(features, dim=1)\n",
"\n",
"print(\"✓ DCT Frequency Analyzer defined\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 6: FFT Spectral Analyzer\n",
"# ============================================================\n",
"\n",
"class FFTSpectralAnalyzer(nn.Module):\n",
" \"\"\"Azimuthally averaged FFT power spectrum.\n",
" Natural images follow 1/f^2 power law. AI images deviate.\"\"\"\n",
" def __init__(self, num_bins=32):\n",
" super().__init__()\n",
" self.num_bins = num_bins\n",
"\n",
" def forward(self, x):\n",
" B, C, H, W = x.shape\n",
" gray = 0.299 * x[:, 0] + 0.587 * x[:, 1] + 0.114 * x[:, 2]\n",
" hann_y = torch.hann_window(H, device=x.device)\n",
" hann_x = torch.hann_window(W, device=x.device)\n",
" window = hann_y.unsqueeze(1) * hann_x.unsqueeze(0)\n",
" gray = gray * window.unsqueeze(0)\n",
"\n",
" fft = torch.fft.fft2(gray)\n",
" fft_shift = torch.fft.fftshift(fft)\n",
" power = torch.abs(fft_shift) ** 2\n",
"\n",
" cy, cx = H // 2, W // 2\n",
" y = torch.arange(H, device=x.device).float() - cy\n",
" xx = torch.arange(W, device=x.device).float() - cx\n",
" yy, xx = torch.meshgrid(y, xx, indexing='ij')\n",
" radius = torch.sqrt(yy**2 + xx**2)\n",
"\n",
" max_radius = min(cy, cx)\n",
" bin_width = max_radius / self.num_bins\n",
"\n",
" features = []\n",
" for i in range(self.num_bins):\n",
" r_lo = i * bin_width\n",
" r_hi = (i + 1) * bin_width\n",
" mask = ((radius >= r_lo) & (radius < r_hi)).float()\n",
" count = mask.sum() + 1e-8\n",
" bin_power = (power * mask.unsqueeze(0)).sum(dim=(-2, -1)) / count\n",
" features.append(bin_power.unsqueeze(1))\n",
"\n",
" radial_spectrum = torch.cat(features, dim=1)\n",
" log_spectrum = torch.log1p(radial_spectrum)\n",
"\n",
" log_freq = torch.log1p(torch.arange(self.num_bins, device=x.device).float() + 1)\n",
" log_freq = log_freq.unsqueeze(0).expand(B, -1)\n",
"\n",
" xm = log_freq - log_freq.mean(dim=1, keepdim=True)\n",
" ym = log_spectrum - log_spectrum.mean(dim=1, keepdim=True)\n",
" slope = (xm * ym).sum(dim=1, keepdim=True) / ((xm**2).sum(dim=1, keepdim=True) + 1e-8)\n",
" intercept = log_spectrum.mean(dim=1, keepdim=True) - slope * log_freq.mean(dim=1, keepdim=True)\n",
"\n",
" predicted = slope * log_freq + intercept\n",
" residuals = log_spectrum - predicted\n",
" residual_std = residuals.std(dim=1, keepdim=True)\n",
" residual_max = residuals.max(dim=1, keepdim=True)[0]\n",
"\n",
" return torch.cat([log_spectrum, slope, intercept, residual_std, residual_max], dim=1)\n",
"\n",
"print(\"✓ FFT Spectral Analyzer defined\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 7: Full Model — FrequencyAwareDetector\n",
"# ============================================================\n",
"\n",
"class FrequencyAwareDetector(nn.Module):\n",
" \"\"\"\n",
" 4-Branch Fusion Detector:\n",
" Branch 1: SwinV2-Tiny (semantic features) → 768d\n",
" Branch 2: SRM filter bank → CNN (forensic residuals) → 256d\n",
" Branch 3: DCT frequency band statistics → 22d\n",
" Branch 4: FFT radial power spectrum → 36d\n",
" Fusion MLP: (768 + 128) → 512 → 128 → 2\n",
" \"\"\"\n",
" def __init__(self, backbone_name=\"microsoft/swinv2-tiny-patch4-window8-256\",\n",
" num_labels=2, dct_patch_size=32, num_freq_bands=8, fft_bins=32):\n",
" super().__init__()\n",
" self.num_labels = num_labels\n",
" self.supports_gradient_checkpointing = True\n",
"\n",
" # Branch 1: SwinV2 backbone\n",
" self.backbone = Swinv2Model.from_pretrained(backbone_name)\n",
" backbone_dim = self.backbone.config.hidden_size # 768\n",
"\n",
" # Branch 2: SRM forensic features\n",
" self.srm = SRMFilterBank()\n",
" self.srm_encoder = nn.Sequential(\n",
" nn.Conv2d(30, 64, 3, stride=2, padding=1), nn.BatchNorm2d(64), nn.GELU(),\n",
" nn.Conv2d(64, 128, 3, stride=2, padding=1), nn.BatchNorm2d(128), nn.GELU(),\n",
" nn.Conv2d(128, 256, 3, stride=2, padding=1), nn.BatchNorm2d(256), nn.GELU(),\n",
" nn.AdaptiveAvgPool2d((1, 1)), nn.Flatten(),\n",
" )\n",
" srm_dim = 256\n",
"\n",
" # Branch 3: DCT\n",
" self.dct_analyzer = DCTFrequencyAnalyzer(patch_size=dct_patch_size, num_freq_bands=num_freq_bands)\n",
" dct_dim = num_freq_bands * 2 + 6 # 22\n",
"\n",
" # Branch 4: FFT\n",
" self.fft_analyzer = FFTSpectralAnalyzer(num_bins=fft_bins)\n",
" fft_dim = fft_bins + 4 # 36\n",
"\n",
" # Frequency projection\n",
" freq_total_dim = srm_dim + dct_dim + fft_dim # 314\n",
" self.freq_proj = nn.Sequential(\n",
" nn.Linear(freq_total_dim, 256), nn.GELU(), nn.Dropout(0.3),\n",
" nn.Linear(256, 128),\n",
" )\n",
"\n",
" # Fusion classifier\n",
" self.classifier = nn.Sequential(\n",
" nn.Linear(backbone_dim + 128, 512), nn.GELU(), nn.Dropout(0.3),\n",
" nn.Linear(512, 128), nn.GELU(), nn.Dropout(0.1),\n",
" nn.Linear(128, num_labels),\n",
" )\n",
"\n",
" self.loss_fn = nn.CrossEntropyLoss(label_smoothing=0.1)\n",
"\n",
" def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None):\n",
" self.backbone.gradient_checkpointing_enable(gradient_checkpointing_kwargs=gradient_checkpointing_kwargs)\n",
"\n",
" def gradient_checkpointing_disable(self):\n",
" self.backbone.gradient_checkpointing_disable()\n",
"\n",
" def forward(self, pixel_values, labels=None, **kwargs):\n",
" semantic_feats = self.backbone(pixel_values=pixel_values).pooler_output\n",
" srm_feats = self.srm_encoder(self.srm(pixel_values))\n",
" dct_feats = self.dct_analyzer(pixel_values)\n",
" fft_feats = self.fft_analyzer(pixel_values)\n",
"\n",
" freq_feats = torch.cat([srm_feats, dct_feats, fft_feats], dim=1)\n",
" freq_proj = self.freq_proj(freq_feats)\n",
"\n",
" fused = torch.cat([semantic_feats, freq_proj], dim=1)\n",
" logits = self.classifier(fused)\n",
"\n",
" loss = None\n",
" if labels is not None:\n",
" loss = self.loss_fn(logits, labels)\n",
"\n",
" return {\"loss\": loss, \"logits\": logits}\n",
"\n",
"\n",
"# Quick test\n",
"print(\"Building model...\")\n",
"model = FrequencyAwareDetector()\n",
"total_params = sum(p.numel() for p in model.parameters())\n",
"print(f\"✓ Model built: {total_params:,} parameters ({total_params/1e6:.1f}M)\")\n",
"del model\n",
"torch.cuda.empty_cache()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 8: Dataset — load from Kaggle local path\n",
"# ============================================================\n",
"\n",
"class AIImageDataset(Dataset):\n",
" \"\"\"\n",
" Loads images from folder structure. Auto-detects:\n",
" - real/ + fake/ (or REAL/ + FAKE/)\n",
" - train/real/ + train/fake/ (nested)\n",
" - any folder with 'real' in name → label 0, others → label 1\n",
" \"\"\"\n",
" def __init__(self, root_path, transform=None, is_train=True):\n",
" self.transform = transform\n",
" self.is_train = is_train\n",
" self.samples = [] # list of (path, label)\n",
"\n",
" root = Path(root_path)\n",
" subdirs = sorted([d for d in root.iterdir() if d.is_dir()])\n",
" subdir_names = [d.name.lower() for d in subdirs]\n",
"\n",
" # Detect structure\n",
" if 'test' in subdir_names and 'train' in subdir_names:\n",
" # Nested: train/real, train/fake, test/real, test/fake\n",
" print(\"Detected train/test split structure\")\n",
" for split_dir in subdirs:\n",
" self._load_class_folders(split_dir)\n",
" else:\n",
" self._load_class_folders(root)\n",
"\n",
" random.shuffle(self.samples)\n",
" labels = [s[1] for s in self.samples]\n",
" print(f\" Total: {len(self.samples)} images\")\n",
" print(f\" Real (0): {labels.count(0)}, Fake (1): {labels.count(1)}\")\n",
"\n",
" def _load_class_folders(self, parent):\n",
" extensions = {'.jpg', '.jpeg', '.png', '.webp', '.bmp', '.tiff'}\n",
" for class_dir in sorted(parent.iterdir()):\n",
" if not class_dir.is_dir():\n",
" continue\n",
" name_lower = class_dir.name.lower()\n",
" # Determine label\n",
" if any(k in name_lower for k in ['real', 'authentic', 'natural', 'photo']):\n",
" label = 0\n",
" elif any(k in name_lower for k in ['fake', 'ai', 'generated', 'synthetic', 'midjourney',\n",
" 'stable', 'dall', 'diffusion']):\n",
" label = 1\n",
" else:\n",
" print(f\" ⚠ Unknown folder '{class_dir.name}' — skipping. \"\n",
" f\"Rename to 'real' or 'fake' to include.\")\n",
" continue\n",
"\n",
" files = [f for f in class_dir.rglob('*') if f.suffix.lower() in extensions]\n",
" self.samples.extend([(str(f), label) for f in files])\n",
" label_name = 'real' if label == 0 else 'fake'\n",
" print(f\" 📁 {class_dir.name}/: {len(files)} files → label={label_name}\")\n",
"\n",
" def __len__(self):\n",
" return len(self.samples)\n",
"\n",
" def __getitem__(self, idx):\n",
" path, label = self.samples[idx]\n",
" try:\n",
" img = Image.open(path).convert('RGB')\n",
" except Exception:\n",
" # Return a black image on error to avoid crashing\n",
" img = Image.new('RGB', (256, 256), (0, 0, 0))\n",
"\n",
" if self.is_train:\n",
" img = self._augment(img)\n",
"\n",
" if self.transform:\n",
" pixel_values = self.transform(img)\n",
" else:\n",
" pixel_values = ToTensor()(img)\n",
"\n",
" return {'pixel_values': pixel_values, 'labels': label}\n",
"\n",
" def _augment(self, img):\n",
" \"\"\"Social media simulation: JPEG compression, blur, downscale.\"\"\"\n",
" if random.random() < 0.10:\n",
" quality = random.randint(30, 95)\n",
" buf = io.BytesIO()\n",
" img.save(buf, format='JPEG', quality=quality)\n",
" buf.seek(0)\n",
" img = Image.open(buf).convert('RGB')\n",
" if random.random() < 0.10:\n",
" img = img.filter(ImageFilter.GaussianBlur(radius=random.uniform(0.1, 2.0)))\n",
" if random.random() < 0.05:\n",
" w, h = img.size\n",
" s = random.uniform(0.5, 0.9)\n",
" img = img.resize((int(w*s), int(h*s)), Image.BILINEAR).resize((w, h), Image.BILINEAR)\n",
" return img\n",
"\n",
"\n",
"# ---------- Load dataset ----------\n",
"print(\"Loading dataset...\")\n",
"IMG_SIZE = 256\n",
"\n",
"train_transform = Compose([\n",
" RandomResizedCrop((IMG_SIZE, IMG_SIZE), scale=(0.8, 1.0)),\n",
" RandomHorizontalFlip(),\n",
" ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1),\n",
" ToTensor(),\n",
" Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),\n",
"])\n",
"\n",
"eval_transform = Compose([\n",
" Resize((IMG_SIZE + 32, IMG_SIZE + 32)),\n",
" CenterCrop((IMG_SIZE, IMG_SIZE)),\n",
" ToTensor(),\n",
" Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),\n",
"])\n",
"\n",
"full_dataset = AIImageDataset(BASE_PATH, transform=None, is_train=False)\n",
"\n",
"# Stratified 90/10 split\n",
"n_total = len(full_dataset)\n",
"n_eval = max(int(0.10 * n_total), 100)\n",
"n_train = n_total - n_eval\n",
"\n",
"generator = torch.Generator().manual_seed(42)\n",
"train_indices, eval_indices = random_split(range(n_total), [n_train, n_eval], generator=generator)\n",
"\n",
"class SubsetWithTransform(Dataset):\n",
" def __init__(self, base_dataset, indices, transform, is_train):\n",
" self.base = base_dataset\n",
" self.indices = list(indices)\n",
" self.transform = transform\n",
" self.is_train = is_train\n",
" def __len__(self):\n",
" return len(self.indices)\n",
" def __getitem__(self, idx):\n",
" real_idx = self.indices[idx]\n",
" path, label = self.base.samples[real_idx]\n",
" try:\n",
" img = Image.open(path).convert('RGB')\n",
" except Exception:\n",
" img = Image.new('RGB', (256, 256), (0, 0, 0))\n",
" if self.is_train:\n",
" img = self.base._augment(img)\n",
" pixel_values = self.transform(img)\n",
" return {'pixel_values': pixel_values, 'labels': label}\n",
"\n",
"train_dataset = SubsetWithTransform(full_dataset, train_indices, train_transform, is_train=True)\n",
"eval_dataset = SubsetWithTransform(full_dataset, eval_indices, eval_transform, is_train=False)\n",
"\n",
"print(f\"\\n✓ Train: {len(train_dataset)}, Eval: {len(eval_dataset)}\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 9: Training Configuration\n",
"# ============================================================\n",
"\n",
"# -------- CONFIGURATION — edit these --------\n",
"HUB_MODEL_ID = \"Reju983/ai-generated-image-detector\" # ← your HF repo\n",
"NUM_EPOCHS = 5\n",
"BATCH_SIZE = 12 # per GPU → effective = 12 * 2 GPUs * 2 accum = 48\n",
"GRAD_ACCUM = 2\n",
"LEARNING_RATE = 2e-5\n",
"OUTPUT_DIR = \"/kaggle/working/ai-image-detector\"\n",
"# -------------------------------------------\n",
"\n",
"print(f\"Effective batch size: {BATCH_SIZE} × {torch.cuda.device_count()} GPUs × {GRAD_ACCUM} accum = \"\n",
" f\"{BATCH_SIZE * torch.cuda.device_count() * GRAD_ACCUM}\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 10: Custom Trainer + Metrics\n",
"# ============================================================\n",
"\n",
"class FreqDetectorTrainer(Trainer):\n",
" def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):\n",
" labels = inputs.pop(\"labels\")\n",
" outputs = model(pixel_values=inputs[\"pixel_values\"], labels=labels)\n",
" loss = outputs[\"loss\"]\n",
" return (loss, outputs) if return_outputs else loss\n",
"\n",
"\n",
"def data_collator(features):\n",
" pixel_values = torch.stack([f[\"pixel_values\"] for f in features])\n",
" labels = torch.tensor([f[\"labels\"] for f in features], dtype=torch.long)\n",
" return {\"pixel_values\": pixel_values, \"labels\": labels}\n",
"\n",
"\n",
"accuracy_metric = evaluate.load(\"accuracy\")\n",
"\n",
"def compute_metrics(eval_pred):\n",
" preds, labels = eval_pred\n",
" if isinstance(preds, dict):\n",
" preds = preds[\"logits\"]\n",
" preds = np.argmax(preds, axis=1)\n",
" acc = accuracy_metric.compute(predictions=preds, references=labels)\n",
" real_mask = labels == 0\n",
" fake_mask = labels == 1\n",
" real_acc = float((preds[real_mask] == labels[real_mask]).mean()) if real_mask.sum() > 0 else 0.0\n",
" fake_acc = float((preds[fake_mask] == labels[fake_mask]).mean()) if fake_mask.sum() > 0 else 0.0\n",
" return {\n",
" \"accuracy\": acc[\"accuracy\"],\n",
" \"real_accuracy\": real_acc,\n",
" \"fake_accuracy\": fake_acc,\n",
" }\n",
"\n",
"print(\"✓ Trainer + Metrics ready\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 11: 🚀 TRAIN!\n",
"# ============================================================\n",
"\n",
"# Init tracking\n",
"trackio.init(project=\"ai-image-detector\", name=\"kaggle-swinv2-srm-dct-fft\")\n",
"\n",
"# Build model fresh\n",
"print(\"Building model...\")\n",
"model = FrequencyAwareDetector(\n",
" backbone_name=\"microsoft/swinv2-tiny-patch4-window8-256\",\n",
" num_labels=2, dct_patch_size=32, num_freq_bands=8, fft_bins=32,\n",
")\n",
"total_params = sum(p.numel() for p in model.parameters())\n",
"print(f\"Parameters: {total_params:,}\")\n",
"\n",
"# Training args — optimized for Kaggle T4 x2\n",
"training_args = TrainingArguments(\n",
" output_dir=OUTPUT_DIR,\n",
" remove_unused_columns=False,\n",
"\n",
" # Schedule\n",
" num_train_epochs=NUM_EPOCHS,\n",
" learning_rate=LEARNING_RATE,\n",
" warmup_ratio=0.1,\n",
" weight_decay=0.01,\n",
" lr_scheduler_type=\"cosine\",\n",
"\n",
" # Batch\n",
" per_device_train_batch_size=BATCH_SIZE,\n",
" per_device_eval_batch_size=BATCH_SIZE,\n",
" gradient_accumulation_steps=GRAD_ACCUM,\n",
"\n",
" # Precision — T4 supports fp16 (NOT bf16)\n",
" fp16=True,\n",
" bf16=False,\n",
"\n",
" # Eval & Save\n",
" eval_strategy=\"epoch\",\n",
" save_strategy=\"epoch\",\n",
" load_best_model_at_end=True,\n",
" metric_for_best_model=\"accuracy\",\n",
" greater_is_better=True,\n",
" save_total_limit=2,\n",
"\n",
" # Logging\n",
" logging_steps=50,\n",
" logging_strategy=\"steps\",\n",
" logging_first_step=True,\n",
" disable_tqdm=False, # keep tqdm in notebooks\n",
" report_to=\"trackio\",\n",
" run_name=\"kaggle-swinv2-srm-dct-fft\",\n",
"\n",
" # Memory\n",
" gradient_checkpointing=True,\n",
" dataloader_num_workers=4,\n",
" dataloader_pin_memory=True,\n",
"\n",
" # Hub\n",
" push_to_hub=True,\n",
" hub_model_id=HUB_MODEL_ID,\n",
" hub_strategy=\"end\",\n",
"\n",
" label_names=[\"labels\"],\n",
")\n",
"\n",
"trainer = FreqDetectorTrainer(\n",
" model=model,\n",
" args=training_args,\n",
" data_collator=data_collator,\n",
" train_dataset=train_dataset,\n",
" eval_dataset=eval_dataset,\n",
" compute_metrics=compute_metrics,\n",
")\n",
"\n",
"print(\"\\n\" + \"=\"*60)\n",
"print(\"Starting training...\")\n",
"print(f\" Train samples: {len(train_dataset)}\")\n",
"print(f\" Eval samples: {len(eval_dataset)}\")\n",
"print(f\" Epochs: {NUM_EPOCHS}\")\n",
"print(f\" Batch: {BATCH_SIZE} x {torch.cuda.device_count()} GPUs x {GRAD_ACCUM} accum\")\n",
"print(f\" LR: {LEARNING_RATE}\")\n",
"print(\"=\"*60 + \"\\n\")\n",
"\n",
"trainer.train()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 12: Evaluate + Push to Hub\n",
"# ============================================================\n",
"\n",
"print(\"\\n\" + \"=\"*60)\n",
"print(\"Final Evaluation:\")\n",
"metrics = trainer.evaluate()\n",
"for k, v in metrics.items():\n",
" print(f\" {k}: {v}\")\n",
"print(\"=\"*60)\n",
"\n",
"# Save model state dict\n",
"save_dir = os.path.join(OUTPUT_DIR, \"final_model\")\n",
"os.makedirs(save_dir, exist_ok=True)\n",
"torch.save(model.state_dict(), os.path.join(save_dir, \"model_state_dict.pt\"))\n",
"print(f\"\\n✓ Model state dict saved to {save_dir}/model_state_dict.pt\")\n",
"\n",
"# Save config\n",
"config = {\n",
" \"architecture\": \"FrequencyAwareDetector\",\n",
" \"backbone_name\": \"microsoft/swinv2-tiny-patch4-window8-256\",\n",
" \"num_labels\": 2, \"dct_patch_size\": 32, \"num_freq_bands\": 8, \"fft_bins\": 32,\n",
" \"id2label\": {\"0\": \"real\", \"1\": \"ai_generated\"},\n",
" \"label2id\": {\"real\": 0, \"ai_generated\": 1},\n",
" \"training_metrics\": metrics,\n",
"}\n",
"with open(os.path.join(save_dir, \"detector_config.json\"), \"w\") as f:\n",
" json.dump(config, f, indent=2)\n",
"\n",
"# Push to Hub\n",
"try:\n",
" print(\"\\nPushing to Hub...\")\n",
" trainer.push_to_hub(\n",
" commit_message=f\"Kaggle training: acc={metrics.get('eval_accuracy', 'N/A')}\",\n",
" tags=[\"image-classification\", \"ai-image-detection\", \"deepfake-detection\",\n",
" \"frequency-analysis\", \"swinv2\", \"srm\", \"dct\", \"fft\"],\n",
" )\n",
"\n",
" from huggingface_hub import HfApi\n",
" api = HfApi()\n",
" api.upload_file(\n",
" path_or_fileobj=os.path.join(save_dir, \"model_state_dict.pt\"),\n",
" path_in_repo=\"model_state_dict.pt\",\n",
" repo_id=HUB_MODEL_ID,\n",
" )\n",
" api.upload_file(\n",
" path_or_fileobj=os.path.join(save_dir, \"detector_config.json\"),\n",
" path_in_repo=\"detector_config.json\",\n",
" repo_id=HUB_MODEL_ID,\n",
" )\n",
" print(f\"\\n✓ Model pushed to https://huggingface.co/{HUB_MODEL_ID}\")\n",
"except Exception as e:\n",
" print(f\"\\n⚠ Hub push failed: {e}\")\n",
" print(f\" Model saved locally at: {save_dir}\")\n",
" print(f\" Download from Kaggle Output tab.\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"trusted": true
},
"outputs": [],
"source": [
"# ============================================================\n",
"# CELL 13: Quick Inference Test\n",
"# ============================================================\n",
"\n",
"model.eval()\n",
"device = next(model.parameters()).device\n",
"\n",
"# Test on a few images from the eval set\n",
"print(\"\\n🔍 Sample predictions:\\n\")\n",
"for i in range(min(10, len(eval_dataset))):\n",
" item = eval_dataset[i]\n",
" pv = item['pixel_values'].unsqueeze(0).to(device)\n",
" true_label = item['labels']\n",
"\n",
" with torch.no_grad():\n",
" out = model(pixel_values=pv)\n",
" probs = torch.softmax(out['logits'], dim=1)\n",
" pred = probs.argmax(dim=1).item()\n",
" conf = probs[0][pred].item()\n",
"\n",
" labels = {0: 'Real', 1: 'AI-Gen'}\n",
" correct = '✓' if pred == true_label else '✗'\n",
" print(f\" {correct} True={labels[true_label]:6s} | Pred={labels[pred]:6s} ({conf:.1%})\")\n",
"\n",
"print(\"\\n🎉 Training complete!\")"
]
}
],
"metadata": {
"kaggle": {
"accelerator": "nvidiaTeslaT4",
"dataSources": [
{
"datasetId": 0,
"sourceId": 0,
"sourceType": "datasetVersion"
}
],
"dockerImageVersionId": 30000,
"isGpuEnabled": true,
"isInternetEnabled": true,
"language": "python",
"sourceType": "notebook"
},
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"name": "python",
"version": "3.10.0"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
|