## Image Comparison: Perturbed Digit Images ### Overview The image presents a side-by-side comparison of two grayscale images, each depicting a handwritten digit "2" with a small perturbation in the top-left corner. The left image is labeled "(a) 4x4 perturbation," while the right image is labeled "(b) 3x3 perturbation." The perturbations are small checkerboard patterns. ### Components/Axes * **Image (a):** * Title: (a) 4x4 perturbation * Content: Grayscale image of a handwritten digit "2" with a 4x4 checkerboard perturbation in the top-left corner. * **Image (b):** * Title: (b) 3x3 perturbation * Content: Grayscale image of a handwritten digit "2" with a 3x3 checkerboard perturbation in the top-left corner. ### Detailed Analysis or ### Content Details * **Image (a) 4x4 perturbation:** * The digit "2" is rendered in grayscale, with varying shades indicating pixel intensity. * The perturbation in the top-left corner consists of a 4x4 grid of alternating dark and light squares, resembling a checkerboard pattern. * The digit is somewhat blurred or pixelated, suggesting a low resolution or the application of a smoothing filter. * **Image (b) 3x3 perturbation:** * Similar to image (a), the digit "2" is rendered in grayscale. * The perturbation in the top-left corner is a smaller 3x3 checkerboard pattern. * The digit also appears blurred or pixelated. ### Key Observations * The primary difference between the two images is the size of the checkerboard perturbation in the top-left corner (4x4 vs. 3x3). * Both images depict the same handwritten digit "2" with similar levels of blurring or pixelation. * The perturbations are subtle and localized, not significantly affecting the overall appearance of the digit. ### Interpretation The image likely demonstrates the effect of small, localized perturbations on an image, possibly in the context of adversarial attacks on image recognition systems. The comparison suggests that even small changes to an image can be noticeable, although in this case, the perturbations do not drastically alter the perceived digit. The experiment may be designed to test the robustness of a model to such perturbations.

## Pixelated Image Comparison: Digit with Localized Perturbations ### Overview The image displays two side-by-side grayscale pixelated images, labeled (a) and (b). Each image features a handwritten digit '1' centrally positioned on a uniform medium gray background, enclosed within a thick black border. The primary difference between the two images is a localized "perturbation" in the top-left corner, which varies in size and specific pixel arrangement. ### Components/Axes No explicit axes, legends, or numerical scales are present in the image. Each sub-image is a self-contained pixel grid. * **Image (a) Label**: "(a) 4x4 perturbation" - This text is positioned above the left pixelated image. * **Image (b) Label**: "(b) 3x3 perturbation" - This text is positioned above the right pixelated image. ### Detailed Analysis Both images are presented as square pixel grids with a thick black border. The background of both images is a uniform medium gray. * **Image (a): 4x4 perturbation** * **Main Subject**: A white, pixelated representation of the handwritten digit '1' is prominently displayed, slightly offset to the right of the center. The digit consists of a primary vertical stroke, which is slightly curved, and a smaller diagonal stroke extending from its top-left, forming the characteristic serif of a handwritten '1'. The pixels forming the digit range from bright white at its core to lighter shades of gray at its edges, indicating a smooth transition or anti-aliasing effect. * **Perturbation**: In the top-left corner of the image, there is a distinct 4x4 grid of pixels that deviates significantly from the medium gray background. This perturbation forms a small, checkerboard-like pattern composed of varying shades of gray, including pixels that are notably darker (approaching black) and pixels that are lighter (approaching white) than the surrounding background. * **Image (b): 3x3 perturbation** * **Main Subject**: This image displays a handwritten digit '1' that is visually identical in shape, position, and pixel intensity to the digit in image (a). It also features the same primary vertical stroke and top-left diagonal serif, rendered in white and light gray pixels against the medium gray background. * **Perturbation**: In the top-left corner, there is a 3x3 grid of pixels, which is the "perturbation" for this image. Similar to image (a), this perturbation forms a small, checkerboard-like pattern of varying gray shades, including darker and lighter pixels that contrast with the medium gray background. The pattern and intensity of the pixels within this 3x3 block appear consistent in style with the 4x4 perturbation, but are confined to a smaller area. ### Key Observations * The core content, the handwritten digit '1', is identical in both images, suggesting it is the baseline or target object. * The background color and the overall framing are consistent across both images. * The only significant difference between the two images lies in the size and specific pixel arrangement of the localized "perturbation" in the top-left corner. * Image (a) features a 4x4 pixel perturbation, while image (b) features a 3x3 pixel perturbation. Both perturbations introduce a small, textured region of varying gray intensities. ### Interpretation These images likely serve to illustrate the impact or presence of localized modifications (perturbations) of different spatial extents on a consistent base image. In the context of technical documents, this type of visualization is commonly used in fields such as: 1. **Machine Learning and Computer Vision**: To demonstrate the effect of small, targeted noise or adversarial examples on an input image. The goal might be to show how a model's classification or feature extraction might be affected by these perturbations, even if the human perception of the digit '1' remains unchanged. 2. **Image Processing**: To compare the visual outcome of applying different sized filters or localized transformations. 3. **Robustness Analysis**: To test the resilience of a system or algorithm to minor, localized changes in input data. The consistent appearance of the digit '1' across both images, despite the differing perturbations, implies that the focus is on how these specific perturbation sizes (4x4 vs. 3x3) might influence an unstated outcome, rather than on altering the primary visual content. The images effectively highlight the precise location and dimensions of these localized changes.

## Image Comparison: Localized Adversarial Perturbations ### Overview The image presents a side-by-side comparison of two grayscale digital images. Both images feature the identical central subject: a pixelated, handwritten digit (resembling a '2') on a uniform gray background, framed by a thick black border. The primary focus of the image is the introduction of localized, square patches of visual noise (perturbations) located in the top-left corners of the gray canvas. ### Components and Layout * **Left Panel Header:** Text reading "(a) 4×4 perturbation" centered above the left image. * **Right Panel Header:** Text reading "(b) 3×3 perturbation" centered above the right image. * **Main Subject (Both Panels):** A white/light-gray curved stroke forming a handwritten numeral '2', positioned centrally. * **Background (Both Panels):** A flat, medium-gray tone. * **Frame (Both Panels):** A thick, solid black border enclosing the gray background and the digit. * **Variable Component:** A small, square grid of altered pixels in the top-left corner, immediately inside the black border. ### Content Details **Panel (a) - Left Image:** * **Label:** (a) 4×4 perturbation * **Perturbation Location:** Top-left corner of the gray background. * **Perturbation Characteristics:** A distinct 4 by 4 grid of pixels (16 pixels total). The pixels within this grid exhibit varying intensities, including dark gray, light gray, and near-white, creating a randomized, checkerboard-like noise pattern. * **Subject Interaction:** The perturbation is spatially isolated from the handwritten digit. There is a clear gap of unaltered gray background between the noise patch and the top of the numeral. **Panel (b) - Right Image:** * **Label:** (b) 3×3 perturbation * **Perturbation Location:** Top-left corner of the gray background, in the exact same anchor position as panel (a). * **Perturbation Characteristics:** A distinct 3 by 3 grid of pixels (9 pixels total). Similar to panel (a), it contains varying intensities of grayscale pixels. It occupies a smaller physical area than the perturbation in panel (a). * **Subject Interaction:** Identical to panel (a), the perturbation is spatially isolated from the handwritten digit. ### Key Observations * **Subject Constancy:** The central handwritten digit and the black border are pixel-perfect matches across both images. The *only* variable is the noise patch in the top-left corner. * **Spatial Isolation:** The noise patches do not obscure, overlap, or touch the primary subject (the digit). * **Dimensionality:** The text explicitly defines the dimensions of the noise patches (4x4 pixels vs. 3x3 pixels), which is visually confirmed by the grid sizes in the respective corners. ### Interpretation * **Context:** This image is a classic visualization from a technical paper in the field of **Adversarial Machine Learning**, specifically focusing on Computer Vision. The handwritten digit is highly characteristic of the MNIST dataset, a standard benchmark in machine learning. * **What the Data Demonstrates:** The image illustrates a "patch attack" or "localized adversarial perturbation." In machine learning, an attacker alters a small number of pixels to intentionally cause a neural network (like a Convolutional Neural Network) to misclassify the image. To a human, the image is obviously still a '2'. To a compromised AI, the specific pattern of noise in the corner forces a wrong prediction (e.g., classifying it as a '7' or '0'). * **Significance of the Comparison (4x4 vs 3x3):** The authors are comparing the size of the adversarial patch. * A **4x4 patch** (16 pixels) gives the attacker more "degrees of freedom" (more variables to optimize) to successfully fool the AI, but alters more of the original image. * A **3x3 patch** (9 pixels) is a more constrained, stealthier attack. It alters fewer pixels (a lower $L_0$ norm), making it harder to detect visually or programmatically, but it may be mathematically harder to find a 3x3 pattern that successfully breaks the model compared to a 4x4 pattern. * **Peircean Investigative Deduction (Reading between the lines):** Because the perturbation is placed in the corner—completely detached from the digit itself—this graphic highlights a specific vulnerability in how the AI processes images. It proves that the AI model is not just looking at the shape of the digit; it is highly sensitive to background/contextual pixels. The model has learned spurious correlations in the background space, allowing an attacker to hijack the classification without ever touching the actual subject of the image.

\n ## Image: Digit Perturbation Comparison ### Overview The image presents a side-by-side comparison of two pixelated images, likely representing a handwritten digit (possibly a '7'). Each image shows the digit after being subjected to a different "perturbation" – a modification of the original pixel values. The goal appears to be to visualize the effect of different perturbation sizes on the digit's appearance. ### Components/Axes The image consists of two sub-images, labeled: * **(a) 4x4 perturbation** – Located on the left. * **(b) 3x3 perturbation** – Located on the right. There are no explicit axes or legends. The images themselves are grayscale pixel arrays. ### Detailed Analysis or Content Details Both images depict a digit that strongly resembles a '7'. The digit is formed by varying shades of gray pixels against a dark background. **Image (a) - 4x4 perturbation:** The digit appears slightly more distorted and blocky compared to image (b). There's a noticeable dark area in the top-left corner, and a few darker pixels scattered along the upper curve of the '7'. The overall shape is still recognizable, but the edges are less smooth. **Image (b) - 3x3 perturbation:** This image shows a smoother, more refined representation of the '7'. The distortion is less pronounced than in image (a). The edges are more defined, and the dark areas are less prominent. The digit appears closer to a clean, unperturbed version. It is impossible to extract numerical data from these images as they are visual representations only. There are no scales or values provided. ### Key Observations * The 3x3 perturbation results in a less distorted image compared to the 4x4 perturbation. * Larger perturbation sizes (4x4) introduce more noticeable artifacts and blockiness. * Both perturbations still allow for the digit to be reasonably identified. ### Interpretation The image demonstrates the impact of perturbation size on the visual quality of a digit image. Perturbation, in this context, likely refers to a process of adding noise or making small changes to the pixel values. A smaller perturbation (3x3) preserves the original shape more effectively, while a larger perturbation (4x4) introduces more significant distortions. This type of visualization is relevant in the field of machine learning, particularly in the context of adversarial attacks. Adversarial attacks involve making small, carefully crafted perturbations to input data (like images) to fool a machine learning model. The image suggests that even relatively small perturbations can alter the appearance of an image, potentially leading to misclassification by a model. The presence of darker pixels in the top-left corner of image (a) could indicate that the perturbation process is not uniform across the image, or that certain areas are more susceptible to distortion. The comparison highlights the trade-off between perturbation size and image quality – larger perturbations can introduce more significant changes, but may also make the image less recognizable. The image does not provide any quantitative data, but it offers a qualitative understanding of how perturbation size affects the visual representation of a digit.

## Comparative Diagram: Image Perturbation Examples ### Overview The image displays two side-by-side grayscale panels, each showing a handwritten digit "1" on a uniform gray background. The panels are labeled to indicate they demonstrate different sizes of localized image perturbations applied to the top-left corner of the digit image. This is likely from a technical context related to image processing, machine learning, or adversarial examples. ### Components/Axes * **Panel Layout:** Two rectangular image panels arranged horizontally. * **Panel Labels:** * Left Panel: `(a) 4×4 perturbation` * Right Panel: `(b) 3×3 perturbation` * **Image Content (Common to Both):** * **Background:** A uniform, medium-gray field. * **Central Figure:** A white, pixelated representation of the handwritten digit "1". The digit has a slight curve at the top and a straight vertical stem. * **Perturbation Region:** A small, square, pixelated pattern located in the top-left corner of each image panel, separate from the digit. * **Perturbation Details:** * **Panel (a):** The perturbation is a 4-pixel by 4-pixel grid. The pattern appears as a checkerboard or alternating sequence of lighter and darker gray pixels. * **Panel (b):** The perturbation is a 3-pixel by 3-pixel grid. The pattern is simpler, with a central darker pixel surrounded by lighter pixels. ### Detailed Analysis * **Spatial Grounding:** The perturbation is consistently placed in the **top-left quadrant** of each image panel, well above and to the left of the digit "1". The digit itself is centered vertically and horizontally within the remaining space. * **Component Isolation:** * **Header Region:** Contains the text labels `(a) 4×4 perturbation` and `(b) 3×3 perturbation` above their respective images. * **Main Image Region:** Contains the gray background, the white digit "1", and the corner perturbation. * **Footer Region:** None present. * **Visual Comparison:** The digit "1" appears identical in shape, position, and pixel intensity in both panels. The sole variable is the size and pattern of the corner perturbation. The 4x4 perturbation in (a) is visibly larger and has a more complex internal pattern than the 3x3 perturbation in (b). ### Key Observations 1. **Isolated Perturbation:** The perturbations are strictly confined to the top-left corner and do not overlap with or alter the central digit "1". 2. **Controlled Variable:** The experiment or demonstration isolates the perturbation size (4x4 vs. 3x3) as the independent variable, while keeping the target digit and background constant. 3. **Pixel-Level Detail:** The images are low-resolution, making the individual pixels of both the digit and the perturbations clearly visible. ### Interpretation This diagram serves as a visual comparison of two different localized image perturbations. In technical fields like machine learning, such perturbations are often used to test the robustness of image classification models (e.g., against adversarial attacks) or to study the effects of noise. * **What it demonstrates:** It shows that a specific, structured noise pattern can be added to an image at different scales (4x4 vs. 3x3 pixels) without affecting the primary subject (the digit "1"). * **Relationship between elements:** The labels directly describe the key difference between the two otherwise nearly identical images. The perturbations are the focal point of the comparison. * **Underlying purpose:** The image likely aims to illustrate a concept where the *size* of a localized disturbance is a parameter of interest. The clarity of the digit despite the perturbations suggests the noise is designed to be subtle or to test model sensitivity to specific, small-scale changes rather than to obscure the image content. The choice of a simple digit "1" provides a clear, unambiguous baseline for such a test.

## Image Analysis: Perturbation Visualization ### Overview The image contains two side-by-side panels labeled **(a) 4×4 perturbation** and **(b) 3×3 perturbation**, each displaying a blurred, curved white line on a gray background. Both panels feature a small checkerboard pattern in the top-left corner. The visual differences between the panels suggest varying degrees of perturbation intensity or resolution. --- ### Components/Axes - **Labels**: - Panel (a): "4×4 perturbation" (top-left text) - Panel (b): "3×3 perturbation" (top-left text) - **Background**: - Uniform gray in both panels. - Checkerboard pattern (black-and-white squares) in the top-left corner of each panel. - **Primary Visual Element**: - A curved, pixelated white line resembling a distorted waveform or path. - No explicit axes, scales, or legends are present. --- ### Detailed Analysis 1. **Checkerboard Pattern**: - Positioned in the top-left corner of both panels. - Likely represents a reference region, mask, or area of interest (e.g., perturbation origin). - No textual labels or annotations within the checkerboard. 2. **Perturbation Lines**: - **Panel (a) 4×4 perturbation**: - Line appears more jagged and fragmented. - Higher spatial resolution (4×4 grid) results in sharper, more defined distortions. - Line curvature is pronounced, with visible "kinks" and irregularities. - **Panel (b) 3×3 perturbation**: - Line is smoother and less fragmented. - Lower spatial resolution (3×3 grid) produces a blurred, averaged effect. - Curvature is less pronounced, with rounded transitions. 3. **Spatial Grounding**: - Checkerboard is consistently placed in the **top-left** of both panels. - Perturbation lines occupy the central and lower regions of each panel. - No overlapping elements or additional annotations. --- ### Key Observations - **Resolution Impact**: The 4×4 perturbation (a) exhibits greater detail and sharper distortions compared to the 3×3 perturbation (b), which appears smoothed due to lower resolution. - **Checkerboard Consistency**: The identical checkerboard pattern in both panels suggests it serves as a fixed reference point, possibly indicating the perturbation's origin or a control region. - **Line Continuity**: Both lines share a similar general shape (curved, diagonal trajectory), but the 4×4 perturbation introduces more localized deviations. --- ### Interpretation The image likely demonstrates the effect of perturbation size on a visual or signal-processing task. The 4×4 perturbation introduces finer, more localized distortions, while the 3×3 perturbation averages these effects, resulting in a smoother output. The checkerboard pattern may represent a masked region or a baseline for comparison. This could relate to applications such as: - **Image compression artifacts**: Comparing how different grid resolutions affect distortion visibility. - **Noise injection in machine learning**: Testing model robustness to varying perturbation intensities. - **Signal processing**: Analyzing how resolution impacts the representation of curved paths or waveforms. The absence of numerical data or explicit axes limits quantitative analysis, but the visual contrast between panels highlights the trade-off between resolution and smoothness in perturbation modeling.