"Snap a Fridge Photo, Get a Meal Plan and Shopping List"… Everyday Security with Eyed Multimodal AI

▲ Photo: AI generated image

Everyone has probably experienced opening the refrigerator door after a tiring day at work, only to sigh at the leftover vegetables and sauces nearing their expiration dates.

Turning on a delivery app feels like a burden on food expenses, and searching for recipes one by one is cumbersome due to matching ingredients. However, recently, by simply turning on a smartphone camera and scanning the inside of a refrigerator, an AI assistant instantly lists the ingredients on the screen and suggests a recipe like "Perilla Seed Stew made in 15 minutes with leftover tofu and mushrooms." As a bonus, it even asks if you'd like to add missing seasonings to your shopping cart. Moving past the era of typing text prompts, 'visual multimodal AI,' which directly sees the world and communicates with humans, has made its way deep into our kitchens and living rooms.

Recently, the AI competition among major big tech companies has rapidly shifted beyond voice and text to 'real-time visual recognition (Vision AI).' Leading tech companies like OpenAI and Google are actively expanding the application of technologies that analyze video frames inputted through cameras in real-time and engage in immediate voice conversations without latency, to general mobile devices. This has evolved beyond mere Optical Character Recognition (OCR) that simply reads text in photos, to a level that comprehensively infers the context and state of objects (freshness, assembly status, broken parts, etc.).

This technological advancement dramatically reduces the fatigue associated with daily chores and overall consumer life. When an unknown error code appears on a washing machine or boiler, instead of sifting through thick manuals, pointing a camera at it can identify the cause and provide emergency solutions. Taking a picture of complex tax bills or hospital receipts can also summarize deductible items and payment deadlines in plain language.

Particularly in the Web3 and fintech industries, attempts are being made to integrate these technologies as practical, everyday tools, such as analyzing spending details from a single photo of a paper receipt to link household ledger token rewards, or visually verifying complex wallet address QR codes to prevent erroneous transfers.

However, the emergence of AI with "eyes" inevitably presents a new challenge: 'visual privacy.' This is because camera lenses capture not only the objects intended by the user but also a wealth of private information in the background.

There is a risk that children's photos, home interiors and structures, bank account copies or contracts carelessly left on a desk, and even the building and unit numbers of apartments visible outside a window could be entirely transmitted to AI servers. If such data is used for model retraining or suffers a cloud breach, a far more three-dimensional invasion of privacy could occur than with text leaks.

Therefore, the fundamental principle for living smartly in the era of visual AI is 'minimizing the shooting frame.' When pointing an object at AI, ensure only the desired subject fills the screen, and adjust the angle to avoid capturing IDs, mail, or family members in the background.

Furthermore, it is crucial to diligently maintain digital hygiene habits, such as restricting camera permissions to 'Allow only while using the app' instead of 'Always allow,' and enabling 'Data learning exclusion (Opt-out)' settings to prevent inputted image data from being reused for corporate AI training.