Hide yourself

The HUMANAID Manifest

1. Never carry your smartphone everywhere without thinking.A smartphone continuously reveals location data, account links, nearby networks, Bluetooth activity, and app behavior. Anyone who wants to remain harder to detect must first understand that a smartphone is often the greatest informant.
2. Only activate functions when you truly need them.Do not leave GPS, Bluetooth, Wi-Fi, NFC, or mobile data permanently enabled. The fewer signals your device sends or receives, the fewer traces it leaves behind.
3. Use as few apps as possible and grant as few permissions as possible.Every app can collect information about location, habits, contacts, or device usage. Install only what is necessary and withdraw permissions that are not essential.
4. Do not use the same digital identity for everything.One fixed account across all services makes it easy to connect behavior, purchases, movements, and communication into one profile. Think carefully before centralizing your life in a single identity.
5. Do not rely entirely on digital payments if you care about privacy.Banking apps, payment cards, and contactless payments reveal time, place, and merchant. Anyone who wants to leave fewer traces must understand that payments can be just as revealing as location data.
6. Move through a camera-filled world with awareness.Streets, stations, shops, apartment buildings, and traffic points are full of cameras. Never assume that only your phone can expose you; your route, your vehicle, your clothing, and your presence can also be recorded.
7. Remember that tracking also happens indoors.Smart buildings, badge readers, Wi-Fi systems, Bluetooth beacons, cameras, smart televisions, video doorbells, and home automation systems can all reveal where you enter, stay, move, and leave. Privacy does not stop at the front door.
8. Do not link all your actions to loyalty cards, reservations, and accounts.Membership cards, subscriptions, tickets, access badges, and online reservations make it easy to reconstruct movement and habits. Convenience is often the enemy of anonymity.
9. Assume that separate traces can later be combined.One camera image, one payment, or one login may say little on its own. But once those fragments are connected, a pattern appears. Protect not only individual actions, but the overall shape of your behavior.
10. Privacy is not a switch but a discipline.Becoming harder to detect does not depend on one trick. It requires consistent choices: sharing less, linking less, automating less, trusting default settings less, and regularly checking which traces you leave behind.

You are always tracable

Adversarial Attacks: When AI sees the wrong thing

Artificial intelligence systems are often presented as objective, fast, and reliable. In practice, many of them are far more fragile than they appear. One of the clearest examples of that fragility is the adversarial attack: a deliberate manipulation of input data that causes a machine-learning model to make a mistake. In computer vision, this can mean that a model misreads a stop sign, fails to detect a face, or confidently classifies an object as something it is not.

At a basic level, adversarial attacks exploit the fact that modern models do not “understand” images the way humans do. A person can still recognize a sign, a face, or a vehicle when part of it is damaged or covered. A neural network, however, may rely on fragile patterns in pixel values and internal feature combinations. Tiny or localized changes can therefore push the model toward the wrong prediction. In digital settings, these changes may be almost invisible. In physical settings, they are often visible, such as a printed patch, occluding accessory, or surface pattern designed to affect classification.

Researchers usually distinguish between digital adversarial examplesand physical adversarial examples. A digital attack modifies the image file directly. A physical attack modifies the real-world scene: a sticker on an object, a pattern on clothing, a patch on an accessory, or another intervention that the camera later captures. Physical attacks are harder to execute consistently because they must remain effective across different viewing angles, lighting conditions, camera sensors, and distances. That is why recent surveys pay special attention to real-world robustness and transferability.

One widely discussed subtype is the adversarial patch. Unlike small imperceptible noise spread across an entire image, a patch attack uses a localized, visible modification that can still mislead a model. This matters because patches are easier to print and deploy in real environments. In the literature, patch attacks have been studied against image classifiers, detectors, and face-recognition systems, precisely because they are practical enough to reveal serious weaknesses in deployed AI.

Face recognition is a particularly sensitive area. Studies and surveys describe physical attacks using accessories, occlusions, light patterns, or adversarial patches aimed at reducing recognition accuracy or causing misidentification. At the same time, ordinary factors such as masks, sunglasses, pose, blur, and illumination already affect facial-recognition performance. In other words, face AI does not only fail under malicious manipulation; it can also degrade under everyday real-world conditions.

This has two important implications. First, adversarial attacks reveal genuine security problems in AI systems used for transport, surveillance, authentication, and safety-critical applications. Second, not every recognition failure is an adversarial attack. Some are simply caused by occlusion, image quality issues, or distribution shift, the gap between training data and the real world. Good technical writing therefore distinguishes among these concepts instead of using adversarialas a catch-all label.

Because of these risks, defenses have become a major area of research. Common strategies include adversarial training, data augmentation, better input-quality checks, multi-view or multi-sensor verification, patch detection, and stronger evaluation under realistic conditions. Surveys consistently argue that no single defense is sufficient on its own. Robustness depends on system design, the quality of training data, operational controls, and continual testing against new attack patterns.

The broader lesson is that AI vision systems are not just software; they are part of a socio-technical environment. Their reliability depends on cameras, lighting, positioning, human behavior, data quality, and security assumptions. Adversarial attacks matter not only because they can be engineered intentionally, but because they expose how brittle many “intelligent” systems still are. If AI is used in important decisions, developers and institutions must evaluate it under real-world stress instead of assuming that good benchmark accuracy means real-world trustworthiness.

Safe ways to reduce your digital and physical discoverability

It is possible to lower your visibility in lawful and privacy-preserving ways without trying to bypass identification systems. In practice, privacy is usually about becoming harder to profile, scrape, and casually locate, not about becoming completely invisible.

A practical privacy-first approach starts with reducing your digital footprint at the source. Delete old accounts you no longer use, remove unused apps, and stop posting identifying details such as your birth date, home address, phone number, travel plans, or routines. The Electronic Frontier Foundation offers useful guidance on managing your digital footprint at EFF’s digital footprint guide.

It is also important to turn off unnecessary location exposure. Disable geotagging in camera and social apps, limit location permissions to only when an app is in use, and avoid posting in real time from your current location. Guidance from the UK Information Commissioner’s Office can be found at ICO online privacy guidance.

Lock down your social profiles wherever possible. Make accounts private, hide contact details, review old posts, remove public follower lists, and ask friends not to tag you in public content without permission. These steps do not make you invisible, but they do lower casual discoverability.

Another important step is to opt out of data brokers and people-search sites where possible. A lot of personal discoverability comes not from your own content, but from third-party aggregation. Reducing the spread of your data across these services can significantly lower your visibility.

You should also use your data rights when applicable. In the UK and EU, for example, the right to erasuremay help in some situations, even though it is not absolute. More information is available at ICO right to erasure guidance.

Another good practice is to separate identities by context. Use different email addresses for shopping, newsletters, professional communication, and personal contacts. Avoid reusing the same username across platforms. Repetition makes cross-platform discovery much easier.

Strong account security also matters. Use unique passwords, a password manager, and two-factor authentication. Old email accounts that are still linked to services can quietly increase your exposure if they are breached or left unmanaged.

You can also minimize searchable public records where lawful and possible. For example, use a business address or PO box for public-facing work instead of a home address when appropriate. Review directory listings, old biographies, and public profiles for outdated details that can be removed or updated.

Physical or analogue visibility matters too. Avoid putting your full name, phone number, or personal address on doorbells, parcels, public noticeboards, or vehicle stickers unless necessary. Shred paperwork that contains personal data, and be careful with printed badges, memberships, and materials that reveal your workplace or daily routine.

The most effective privacy measures are usually simple and consistent. Think in layers rather than absolutes: fewer public identifiers, fewer linked accounts, less routine oversharing, and more deliberate control over the data you leave behind.

Key concepts

• Adversarial attack: a deliberate manipulation intended to make an AI system misclassify or misinterpret input.
• Adversarial example: the altered input itself.
• Adversarial patch: a localized visible modification, often studied in physical-world scenarios.
• Occlusion: a case where part of an object or face is blocked, without necessarily being adversarial.
• Distribution shift: the difference between training data and real-world conditions.

Practical privacy checklist

1. Delete unused accounts and apps.
2. Remove personal details from public profiles.
3. Disable unnecessary location sharing and geotagging.
4. Make social media accounts private where possible.
5. Use separate email addresses for different contexts.
6. Enable two-factor authentication and use unique passwords.
7. Opt out of data brokers and people-search sites.
8. Review public directory listings and outdated biographies.
9. Avoid printing unnecessary personal information on physical materials.
10. Think in layers and reduce visibility step by step.