{"notes":[{"id":"face-antispoofing/iadg","category":"face-antispoofing","legacyPath":"face-antispoofing/2304-iadg","legacySlug":"2304-iadg","legacyPaths":[],"title":"IADG","subtitle":"A Monologue of Styles","author":"Z. Yuan","paperDate":"2023-04-12","publishedOn":"2025-05-06","paperTitle":"Instance-Aware Domain Generalization for Face Anti-Spoofing","paperUrl":"https://arxiv.org/abs/2304.05640","bibliography":{"title":"Instance-Aware Domain Generalization for Face Anti-Spoofing","authors":["Qianyu Zhou","Ke-Yue Zhang","Taiping Yao","Xuequan Lu","Ran Yi","Shouhong Ding","Lizhuang Ma"],"url":"https://arxiv.org/abs/2304.05640","pdfUrl":"https://arxiv.org/pdf/2304.05640","codeUrl":null,"preprint":{"id":"2304.05640","submittedOn":"2023-04-12","url":"https://arxiv.org/abs/2304.05640"},"publication":{"venue":"CVPR","year":2023,"kind":"conference","url":"https://arxiv.org/abs/2304.05640"},"citation":"@misc{faceantispoofingiadg2023,\n  title = {{Instance-Aware Domain Generalization for Face Anti-Spoofing}},\n  author = {Qianyu Zhou and Ke-Yue Zhang and Taiping Yao and Xuequan Lu and Ran Yi and Shouhong Ding and Lizhuang Ma},\n  year = {2023},\n  eprint = {2304.05640},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2304.05640}\n}\n","citationPath":"/content/papers/citations/face-antispoofing/iadg.b4743b4893ce.bib"},"thumbnail":{"src":"/content/papers/assets/502e6feaa5147fdd4a87418e8d09a327b4d060b7d3e2881dba2bca5e518d4857.jpg","width":1778,"height":580},"locale":"en","path":"/content/papers/en/face-antispoofing/iadg.64fb7646eb41.md","headings":[{"id":"defining-the-problem","text":"Defining the Problem","depth":2},{"id":"solving-the-problem","text":"Solving the Problem","depth":2},{"id":"dynamic-kernel-generator","text":"Dynamic Kernel Generator","depth":3},{"id":"categorical-style-assembly","text":"Categorical Style Assembly","depth":3},{"id":"asymmetric-instance-adaptive-whitening","text":"Asymmetric Instance Adaptive Whitening","depth":3},{"id":"overall-training-and-optimization","text":"Overall Training and Optimization","depth":3},{"id":"discussion","text":"Discussion","depth":2},{"id":"comparison-with-existing-methods","text":"Comparison with Existing Methods","depth":3},{"id":"ablation-experiments","text":"Ablation Experiments","depth":3},{"id":"visualization-analysis","text":"Visualization Analysis","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"face-antispoofing/ssdg","relation":"comparison","topic":"Domain generalization for face anti-spoofing"},{"id":"face-antispoofing/ssan","relation":"comparison","topic":"Domain generalization for face anti-spoofing"},{"id":"face-antispoofing/sa-fas","relation":"comparison","topic":"Domain generalization for face anti-spoofing"}]},{"id":"face-antispoofing/ssdg","category":"face-antispoofing","legacyPath":"face-antispoofing/2004-ssdg","legacySlug":"2004-ssdg","legacyPaths":[],"title":"SSDG","subtitle":"Stable Realness","author":"Z. Yuan","paperDate":"2020-04-29","publishedOn":"2025-04-18","paperTitle":"Single-Side Domain Generalization for Face Anti-Spoofing","paperUrl":"https://arxiv.org/abs/2004.14043","bibliography":{"title":"Single-Side Domain Generalization for Face Anti-Spoofing","authors":["Yunpei Jia","Jie Zhang","Shiguang Shan","Xilin Chen"],"url":"https://arxiv.org/abs/2004.14043","pdfUrl":"https://arxiv.org/pdf/2004.14043","codeUrl":null,"preprint":{"id":"2004.14043","submittedOn":"2020-04-29","url":"https://arxiv.org/abs/2004.14043"},"publication":{"venue":"CVPR","year":2020,"kind":"conference","url":"https://doi.org/10.1109/cvpr42600.2020.00851"},"citation":"@misc{faceantispoofingssdg2020,\n  title = {{Single-Side Domain Generalization for Face Anti-Spoofing}},\n  author = {Yunpei Jia and Jie Zhang and Shiguang Shan and Xilin Chen},\n  year = {2020},\n  eprint = {2004.14043},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2004.14043}\n}\n","citationPath":"/content/papers/citations/face-antispoofing/ssdg.18ad604217a9.bib"},"thumbnail":{"src":"/content/papers/assets/3ec4bd961265b910a76f4573bbeb850168d7c706501e6f4a51e8c02a71c72b26.jpg","width":1490,"height":756},"locale":"en","path":"/content/papers/en/face-antispoofing/ssdg.03326184e995.md","headings":[{"id":"problem-definition","text":"Problem Definition","depth":2},{"id":"solution","text":"Solution","depth":2},{"id":"single-side-adversarial-learning","text":"Single-Side Adversarial Learning","depth":3},{"id":"asymmetric-triplet-loss","text":"Asymmetric Triplet Loss","depth":3},{"id":"feature-and-weight-normalization","text":"Feature and Weight Normalization","depth":3},{"id":"additional-supplement","text":"Additional Supplement","depth":2},{"id":"discussion","text":"Discussion","depth":2},{"id":"dataset-testing-setup","text":"Dataset Testing Setup","depth":3},{"id":"comparison-with-other-methods","text":"Comparison with Other Methods","depth":3},{"id":"ablation-experiment","text":"Ablation Experiment","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"face-antispoofing/ssan","relation":"comparison","topic":"Domain generalization for face anti-spoofing"},{"id":"face-antispoofing/sa-fas","relation":"comparison","topic":"Domain generalization for face anti-spoofing"},{"id":"face-antispoofing/iadg","relation":"comparison","topic":"Domain generalization for face anti-spoofing"}]},{"id":"face-antispoofing/ssan","category":"face-antispoofing","legacyPath":"face-antispoofing/2203-ssan","legacySlug":"2203-ssan","legacyPaths":[],"title":"SSAN","subtitle":"The Shadow of Style","author":"Z. Yuan","paperDate":"2022-03-10","publishedOn":"2025-04-30","paperTitle":"Domain Generalization via Shuffled Style Assembly for Face Anti-Spoofing","paperUrl":"https://arxiv.org/abs/2203.05340","bibliography":{"title":"Domain Generalization via Shuffled Style Assembly for Face Anti-Spoofing","authors":["Zhuo Wang","Zezheng Wang","Zitong Yu","Weihong Deng","Jiahong Li","Tingting Gao","Zhongyuan Wang"],"url":"https://arxiv.org/abs/2203.05340","pdfUrl":"https://arxiv.org/pdf/2203.05340","codeUrl":null,"preprint":{"id":"2203.05340","submittedOn":"2022-03-10","url":"https://arxiv.org/abs/2203.05340"},"publication":{"venue":"CVPR","year":2022,"kind":"conference","url":"https://arxiv.org/abs/2203.05340"},"citation":"@misc{faceantispoofingssan2022,\n  title = {{Domain Generalization via Shuffled Style Assembly for Face Anti-Spoofing}},\n  author = {Zhuo Wang and Zezheng Wang and Zitong Yu and Weihong Deng and Jiahong Li and Tingting Gao and Zhongyuan Wang},\n  year = {2022},\n  eprint = {2203.05340},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2203.05340}\n}\n","citationPath":"/content/papers/citations/face-antispoofing/ssan.65b52ab509cf.bib"},"thumbnail":{"src":"/content/papers/assets/31e7f21274d546d558932f8110ce87e7599aa0e046f3fecdea4927eab53dc1cb.jpg","width":1372,"height":786},"locale":"en","path":"/content/papers/en/face-antispoofing/ssan.2e557af140cf.md","headings":[{"id":"defining-the-problem","text":"Defining the Problem","depth":2},{"id":"solving-the-problem","text":"Solving the Problem","depth":2},{"id":"dual-stream-disentangling-of-content-and-style","text":"Dual-Stream Disentangling of Content and Style","depth":3},{"id":"style-reassembly-and-shuffling","text":"Style Reassembly and Shuffling","depth":3},{"id":"contrastive-learning-in-style","text":"Contrastive Learning in Style","depth":3},{"id":"overall-loss","text":"Overall Loss","depth":3},{"id":"discussion","text":"Discussion","depth":2},{"id":"comparison-with-other-methods","text":"Comparison with Other Methods","depth":3},{"id":"large-scale-testing","text":"Large-Scale Testing","depth":3},{"id":"feature-visualization","text":"Feature Visualization","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"face-antispoofing/ssdg","relation":"comparison","topic":"Domain generalization for face anti-spoofing"},{"id":"face-antispoofing/sa-fas","relation":"comparison","topic":"Domain generalization for face anti-spoofing"},{"id":"face-antispoofing/iadg","relation":"comparison","topic":"Domain generalization for face anti-spoofing"}]},{"id":"face-antispoofing/sa-fas","category":"face-antispoofing","legacyPath":"face-antispoofing/2303-sa-fas","legacySlug":"2303-sa-fas","legacyPaths":[],"title":"SA-FAS","subtitle":"The Law of the Hyperplane","author":"Z. Yuan","paperDate":"2023-03-23","publishedOn":"2025-04-18","paperTitle":"Rethinking Domain Generalization for Face Anti-spoofing: Separability and Alignment","paperUrl":"https://arxiv.org/abs/2303.13662","bibliography":{"title":"Rethinking Domain Generalization for Face Anti-spoofing: Separability and Alignment","authors":["Yiyou Sun","Yaojie Liu","Xiaoming Liu","Yixuan Li","Wen-Sheng Chu"],"url":"https://arxiv.org/abs/2303.13662","pdfUrl":"https://arxiv.org/pdf/2303.13662","codeUrl":null,"preprint":{"id":"2303.13662","submittedOn":"2023-03-23","url":"https://arxiv.org/abs/2303.13662"},"publication":{"venue":"CVPR","year":2023,"kind":"conference","url":"https://arxiv.org/abs/2303.13662"},"citation":"@misc{faceantispoofingsafas2023,\n  title = {{Rethinking Domain Generalization for Face Anti-spoofing: Separability and Alignment}},\n  author = {Yiyou Sun and Yaojie Liu and Xiaoming Liu and Yixuan Li and Wen-Sheng Chu},\n  year = {2023},\n  eprint = {2303.13662},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2303.13662}\n}\n","citationPath":"/content/papers/citations/face-antispoofing/sa-fas.a0e5f70569ad.bib"},"thumbnail":{"src":"/content/papers/assets/c5d5b32f66a70c73d21f58ebcfc0fdc186c0dac32b554fd61aa098f945fa6536.jpg","width":1518,"height":394},"locale":"en","path":"/content/papers/en/face-antispoofing/sa-fas.77a3a739b355.md","headings":[{"id":"defining-the-problem","text":"Defining the Problem","depth":2},{"id":"solving-the-problem","text":"Solving the Problem","depth":2},{"id":"separability","text":"Separability","depth":3},{"id":"alignment","text":"Alignment","depth":3},{"id":"projected-gradient-irm-pg-irm","text":"Projected Gradient IRM (PG-IRM)","depth":3},{"id":"α-adjacency-is-the-key","text":"α-adjacency is the Key","depth":3},{"id":"discussion","text":"Discussion","depth":2},{"id":"comparison-with-other-methods","text":"Comparison with Other Methods","depth":3},{"id":"evaluation-metric-analysis","text":"Evaluation Metric Analysis","depth":3},{"id":"feature-distribution","text":"Feature Distribution","depth":3},{"id":"comparison-with-dann-and-ssdg","text":"Comparison with DANN and SSDG","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"face-antispoofing/ssdg","relation":"comparison","topic":"Domain generalization for face anti-spoofing"},{"id":"face-antispoofing/ssan","relation":"comparison","topic":"Domain generalization for face anti-spoofing"},{"id":"face-antispoofing/iadg","relation":"comparison","topic":"Domain generalization for face anti-spoofing"}]}],"text":"Over the past decade, researchers have tried every possible method to draw the boundary between the real and the fake:\n\n- Using handcrafted features like LBP, SIFT, and HOG to recognize textures and details;\n- Exploring anomalies in the YCrCb color space and Fourier spectrum;\n- Analyzing subtle differences in time-series, attempting to find evidence of a real face from blink, micro-movements, and blood flow;\n\nThen, the whole field passed the baton to CNN, using classifiers, depth maps, reflectance maps, and rPPG for auxiliary supervision.\n\nHowever, these methods often experience a significant drop in performance when the scene, device, or attack method changes, making them difficult to deploy.\n\n## Defining the Problem\n\nTo make models adaptable to unknown domains, researchers introduced the concepts of **Domain Adaptation (DA)** and **Domain Generalization (DG)**.\n\nIn particular, in DG, adversarial learning and meta-learning became mainstream choices, aiming to learn a representation that says, \"No matter where you come from, I can spot you.\"\n\nBut as this path progressed, its limitations began to emerge:\n\n- Many DG methods use explicit domain labels, although methods such as D²AM already avoid manually defined domain labels;\n- Dataset names, device models, and other human-defined groups are useful organizational labels, but may not capture fine variations in lighting and background.\n\nYou say \"this is Replay,\" but Replay contains multiple lighting conditions and backgrounds; you say \"this is Oulu,\" but no one can specify how many style subgroups it contains.\n\nEven with labels, these methods only align features at the distribution level but fail to control for certain feature channels that still secretly retain style traces. These seemingly harmless signals often become blind spots for the model when faced with new samples.\n\nHow can we distinguish live faces from attacks without domain labels? Live/spoof labels are still required; this is not unsupervised learning.\n\nIs there another boundary between real and fake?\n\n## Solving the Problem\n\n\n\n![model arch](/content/papers/assets/502e6feaa5147fdd4a87418e8d09a327b4d060b7d3e2881dba2bca5e518d4857.jpg)\n\n\n\nTo solve this problem, the authors propose a new architecture, called:\n\n- **Instance-Aware Domain Generalization (IADG)**.\n\nThe goal is to enable the model to learn \"instance-level\" features without relying on domain labels, and to maintain stable performance in unknown domains.\n\nIt primarily consists of three core components:\n\n- **Asymmetric Instance Adaptive Whitening (AIAW)**\n- **Categorical Style Assembly (CSA)**\n- **Dynamic Kernel Generator (DKG)**\n\nLet's take a closer look at the design of each module in order.\n\n### Dynamic Kernel Generator\n\n\n\n![dkg](/content/papers/assets/1567cec0ee0e2956aa90ce3079f2977de2cd9be64ef415f7f581cfedc156e304.jpg)\n\n\n\nSince samples from multiple source domains often have high diversity, a single static filter cannot meet the feature needs of all instances. To address this, the authors designed **DKG** to automatically generate \"instance-adaptive\" dynamic filters, working alongside static filters to capture more comprehensive individual perception features, aiding subsequent cross-domain generalization.\n\nThe structure of DKG consists of two parts: \"static convolution branch\" and \"dynamic convolution branch\":\n\n- The static branch shares its convolution parameters across samples; those parameters are still updated during training;\n- The dynamic branch generates convolution kernel parameters dynamically based on the features of each sample.\n\nAs shown in the figure above, let $X_i$ and $F_i$ represent the input and output features of the $i$-th sample in DKG.\n\nDuring training, both branches learn synchronously, and the process is as follows:\n\n1. The input feature $X_i$ is split along the channel dimension into $\\hat{X}_i$ and $\\tilde{X}_i$;\n2. **Static convolution branch**: $\\tilde{X}_i$ is processed by a static convolution filter $f_{\\theta_s}$ to produce $\\tilde{Z}_i$;\n3. **Dynamic convolution branch**: $\\hat{X}_i$ is first passed through global average pooling and a convolution block $f_{\\theta_1}$, generating a dynamic convolution kernel $W_i$ related to this instance. Then, this dynamic kernel is used to convolve $\\hat{X}_i$ to obtain $\\hat{Z}_i$.\n\nFinally, $\\tilde{Z}_i$ and $\\hat{Z}_i$ are concatenated along channels and passed through the convolution block $f_{\\theta_2}$ to obtain $F_i$. This lets the static and dynamic filters work together for each instance.\n\n### Categorical Style Assembly\n\n\n\n![dkg](/content/papers/assets/fdf9c83f79af35f26daebf8a8300dbae00cfe1eada6810273152438607e58d74.jpg)\n\n\n\nCSA generates more diverse instance-level style changes. It improves on earlier augmentation strategies that randomly swap or mix source styles; AdaIN itself neither requires random style selection nor determines class labels.\n\n\n> [!TIP]\n> \n> Let’s pause for a moment, as some readers may not be familiar with AdaIN. Here’s a brief explanation:\n> \n> **Adaptive Instance Normalization (AdaIN)** is a commonly used feature alignment method in style transfer, originally applied in image style transfer tasks. Its core idea is:\n> \n> > **Adjust the mean and standard deviation of content features to match the statistical properties of the target style**, achieving style injection and transformation.\n> \n> Specifically, for content feature $x$ and style feature $y$, the AdaIN computation formula is:\n> \n> $$\n> \\text{AdaIN}(x, y) = \\sigma(y) \\cdot \\left( \\frac{x - \\mu(x)}{\\sigma(x)} \\right) + \\mu(y)\n> $$\n> \n> Where:\n> \n> - $\\mu(x), \\sigma(x)$ are the spatial mean and standard deviation of each channel of each sample;\n> - $\\mu(y), \\sigma(y)$ are the statistical properties of the style features.\n> \n> Replacing channel statistics aims to change style while preserving content. Implementations also need a small stabilizer for near-zero standard deviations.\n> \n> Due to its simplicity and efficiency, AdaIN has been widely used in style transfer and cross-domain generalization tasks, especially in applications that require style alignment or de-stylization.\n> \n\n\nCSA makes two changes to how styles are selected for augmentation:\n\n1. **Diversity Consideration**: FPS iteratively selects styles far from those already selected, building $L$ basis styles and refreshing them each epoch. This encourages diversity, including less common styles, without guaranteeing coverage of every unseen distribution.\n\n2. **Separate Treatment of Real and Fake**: Styles are assembled within the same live/spoof class to reduce the risk of changing discriminative cues. Preserving the label is the design goal, not a general mathematical guarantee.\n\nEach class contributes $L$ basis styles with stored channel mean and variance information. The assembly equations use means $\\mu_{\\mathrm{base}}^c$ and standard deviations $\\sigma_{\\mathrm{base}}^c$; $\\sigma$ must not be confused with variance $\\sigma^2$.\n\nFor class $c$, nonnegative weights summing to one are sampled from a Dirichlet distribution whose concentration parameters are all $1/L$. They combine basis means and standard deviations into $\\mu_{\\mathrm{aug}}$ and $\\sigma_{\\mathrm{aug}}$, then transform a same-class sample as follows:\n\n$$\nF_{\\mathrm{aug}}\n= \\sigma_{\\mathrm{aug}}\n\\Bigl(\n\\frac{F_{\\mathrm{org}} - \\mu(F_{\\mathrm{org}})}{\\sigma(F_{\\mathrm{org}})}\n\\Bigr)\n+ \\mu_{\\mathrm{aug}}\n$$\n\nThis expands style diversity while keeping the sample's original live/spoof label as the training target; the original/augmented feature pairs then help identify correlations sensitive to style changes.\n\n### Asymmetric Instance Adaptive Whitening\n\n\n\n![dkg](/content/papers/assets/9f5441daf4319599c3a549d40b9a1a5c3e6709fee396139bbd29826eca34b175.jpg)\n\n\n\nIn order to \"precisely align\" the features of each sample, the author further focuses on the correlation between feature channels.\n\nPrevious studies have pointed out that the covariance matrix of features often contains style-related information. If whitening is applied to it, it can weaken some domain-specific styles.\n\nHowever, simply applying existing **instance whitening** methods may result in the removal of \"useful discriminative features,\" which could impair the performance of Face Anti-Spoofing (FAS).\n\nThe asymmetric design encourages compact live features while leaving more room for differences among attacks. Its actual procedure is:\n\n1. Apply IN separately to original and CSA-augmented features, flatten each $C\\times H\\times W$ map into $C\\times HW$, and compute $\\Sigma=FF^\\top/(HW)$.\n2. Compute the elementwise variance between the two covariance matrices for each sample, then average across samples to form $V$. Large entries indicate sensitivity to the style perturbation.\n3. Rank entries to construct a binary mask $M(k_c)$, keeping only the strict upper triangle. Select a larger fraction for live samples, $k_r$, than for attacks, $k_s$.\n4. Penalize the selected entries $\\Sigma_t\\odot M(k_c)$ in both original and augmented branches, pushing them toward zero.\n\nAIAW is thus a selective decorrelation loss, not a layer that multiplies all features by an inverse covariance square root. The following whitening explanation provides background.\n\n\n> [!TIP]\n> \n> **Whitening** is a method used to decorrelate and normalize data features, commonly employed in statistics and deep learning as part of feature preprocessing. The goal is to transform an input feature $x \\in \\mathbb{R}^d$ into a vector $\\hat{x}$ that satisfies the following properties:\n> \n> - Zero mean: $\\mathbb{E}[\\hat{x}] = 0$\n> - Covariance equal to the identity matrix: $\\text{Cov}(\\hat{x}) = I$\n> \n> A common whitening formula is as follows:\n> \n> $$\n> \\hat{x} = W (x - \\mu)\n> $$\n> \n> Here $\\mu$ is the mean. For positive-definite covariance $\\Sigma=U\\Lambda U^\\top$, the following gives ZCA whitening; PCA whitening omits the final $U^\\top$:\n> \n> $$\n> W = \\Sigma^{-\\frac{1}{2}} = U \\Lambda^{-\\frac{1}{2}} U^\\top\n> $$\n> \n> Under ideal conditions this removes correlation and equalizes scale. Singular or unstable covariance estimates require regularization or dimensionality reduction; adding a stabilizer also makes the resulting covariance only approximately identity.\n> \n> Whitening addresses cross-channel correlation, whereas AdaIN primarily replaces per-channel means and standard deviations. AIAW selectively constrains correlations to reduce the loss of discriminative information.\n> \n\n\n### Overall Training and Optimization\n\nTo achieve stable performance in unknown domains, the classifier must be able to correctly distinguish between real and spoofed samples under various \"style disturbances.\"\n\nTherefore, for the same input sample $X_i$, the author supervises the binary classification loss $\\mathcal{L}_{\\mathrm{Cls}}$ using both the original feature $F_{\\mathrm{org}}^i$ and the augmented feature $F_{\\mathrm{aug}}^i$:\n\n$$\n\\mathcal{L}_{\\mathrm{Cls}}\n=-\\sum_i\\sum_{t\\in\\{\\mathrm{org},\\mathrm{aug}\\}}\\sum_{c\\in\\{\\mathrm{live},\\mathrm{spoof}\\}}\nY_{ic}\\log p_{ic}^{t}\n$$\n\nThis expands the paper’s shorthand into two-class cross-entropy: $Y_{ic}$ is a one-hot label, $p_{ic}^{t}$ is branch $t$’s probability for class $c$, and $i$ ranges over source training samples. A single-sigmoid formulation requires both $y\\log p$ and $(1-y)\\log(1-p)$ terms.\n\nAdditionally, the author also references previous research and uses a depth estimator for additional supervision: outputting a depth map for real faces and a zero map for spoofed faces.\n\nThe depth loss function $\\mathcal{L}_{\\mathrm{Dep}}$ is:\n\n$$\n\\mathcal{L}_{\\mathrm{Dep}}\n=\\sum_i\\sum_{t\\in\\{\\mathrm{org},\\mathrm{aug}\\}}\n\\left\\|\\mathrm{Dep}(F_t^i)-Y_{\\mathrm{dep}}^i\\right\\|_2^2.\n$$\n\n$Y_{\\mathrm{dep}}^i$ is a supervision target: PRNet pseudo-depth for live faces and an all-zero map for attacks. It is not sensor-measured depth with the all-zero attack map serving as a training convention.\n\nCombining the above, the final total loss function $\\mathcal{L}_{\\mathrm{total}}$ is:\n\n$$\n\\mathcal{L}_{\\mathrm{total}}\n= \\mathcal{L}_{\\mathrm{Cls}}\n+ \\lambda \\mathcal{L}_{\\mathrm{Dep}}\n+ \\mathcal{L}_{\\mathrm{AIAW}}.\n$$\n\nDuring the inference phase, only the original feature branch is used, and no augmentation operations are performed; the AIAW loss is only active during training.\n\nDKG, CSA, and AIAW reduce sensitivity to style at the instance level. Generalization to a new environment still requires evaluation; the following results apply to the reported datasets and protocols.\n\n## Discussion\n\nAs usual, this section follows the traditional **Leave-One-Out (LOO)** experimental setup, followed by ablation experiments to analyze the contributions of each module.\n\n### Comparison with Existing Methods\n\n\n\n![comparison](/content/papers/assets/1b846375ec3c7db55e1d2feba126800f77d17b56933fea3982598a40c5570faa.jpg)\n\n\n\nTable 1 compares conventional FAS and DG methods. Domain-label granularity motivates the paper, but the table alone does not attribute every performance difference to those labels.\n\nIADG has the lowest reported HTER in all four directions and the highest AUC in three. On OCM→I, its 94.50% AUC is slightly below SSDG-M’s 94.61% and SSAN’s 94.58%. It avoids domain labels, while still requiring live/spoof supervision.\n\n### Ablation Experiments\n\n\n\n![ablation](/content/papers/assets/d8ea7662f4224ae877db6ac7bddb5c1203e4d02496e6bf856013d3d5c0518000.jpg)\n\n\n\nAll ablations below use I&C&M→O; the values are not averages over four directions:\n\n- Using only the backbone (baseline) achieves 19.75% HTER and 87.46% AUC.\n- Adding **DKG** improves the result to 16.94% / 90.14%.\n- Further adding **CSA** advances the performance to 12.50% / 93.62%.\n- Finally, introducing **AIAW** boosts the performance to 8.86% / 97.14%.\n\nIn this cumulative ablation, HTER falls and AUC rises at each addition. The table does not test every separate combination of the three modules.\n\nNext, let's look at the comparison of different Instance Whitening methods:\n\n\n\n![ablation](/content/papers/assets/e395c83b7a190f93e5249165b870ce9b6677d96db8a39f1b6ef78283e0bc6cae.jpg)\n\n\n\nThe experimental results show that methods like IW and GIW treat all covariance elements equally, which can inadvertently remove key discriminative features. While ISW shows some improvement, it still ignores the distributional differences between real and spoofed faces.\n\nAIAW uses different selection fractions for live and spoof samples and constrains both branches. The best tested ratio in Table 5 is $k_r:k_s=1:0.2$, with actual settings $k_r=0.3\\%$ and $k_s=0.06\\%$. These are not 100% and 20% of covariance entries.\n\n### Visualization Analysis\n\n\n\n![visualization](/content/papers/assets/c7f500d1d57434520daadaa800d1434f72a351243c023dae6b2e13741f6d0d02.jpg)\n\n\n\nFinally, to better understand how IADG aligns features, we look at the t-SNE projection.\n\nIn the t-SNE plots, the domain-aware method separates source live and spoof samples, but target live and spoof samples overlap more. With IADG, the target groups appear easier to separate. The figure does not draw a classifier decision boundary.\n\nThe more compact live features agree qualitatively with the design motivation.\n\n## Conclusion\n\nFor face anti-spoofing, drawing the boundary between real and fake is only the start; the harder challenge is handling variation.\n\nThese variations not only arise from the evolution of attack methods but also from the fragmentation of the data itself:\n\n> **Different lighting conditions, devices, backgrounds, cultures, and shooting moments, each piece of data could be its own domain.**\n\nWe once thought domain labels could control everything, but it turns out that more and more differences are hidden in those unclassifiable details. Rather than chasing uniform labels, it is better to extract the invariant signals from the style noise of each face.\n\nFrom aligning the overall picture to understanding the individual; from classifying samples to dissecting style; perhaps we still don’t fully understand the essence of style, but IADG at least offers us a way to \"coexist with style.\"\n","imageSizes":{"/content/papers/assets/502e6feaa5147fdd4a87418e8d09a327b4d060b7d3e2881dba2bca5e518d4857.jpg":{"width":1778,"height":580},"/content/papers/assets/1567cec0ee0e2956aa90ce3079f2977de2cd9be64ef415f7f581cfedc156e304.jpg":{"width":1224,"height":596},"/content/papers/assets/fdf9c83f79af35f26daebf8a8300dbae00cfe1eada6810273152438607e58d74.jpg":{"width":612,"height":248},"/content/papers/assets/9f5441daf4319599c3a549d40b9a1a5c3e6709fee396139bbd29826eca34b175.jpg":{"width":612,"height":220},"/content/papers/assets/1b846375ec3c7db55e1d2feba126800f77d17b56933fea3982598a40c5570faa.jpg":{"width":1362,"height":624},"/content/papers/assets/d8ea7662f4224ae877db6ac7bddb5c1203e4d02496e6bf856013d3d5c0518000.jpg":{"width":1208,"height":368},"/content/papers/assets/e395c83b7a190f93e5249165b870ce9b6677d96db8a39f1b6ef78283e0bc6cae.jpg":{"width":908,"height":376},"/content/papers/assets/c7f500d1d57434520daadaa800d1434f72a351243c023dae6b2e13741f6d0d02.jpg":{"width":1224,"height":560}}}
