Ahmet Semseddin Gultekin

Posted on Sep 9 AI-assisted

The Security of Smart Shopping Carts

#iot #networking #privacy #security

Introduction

In recent years, technologies that let shoppers skip the checkout line in large supermarkets have been spreading fast. Products like Amazon's Dash Cart, Caper AI, Veeve, and Shopic pack RFID (Radio Frequency Identification) readers, cameras, weight sensors, and cloud connectivity into a single cart, promising to eliminate the checkout queue entirely.

In this post, I'll cover how the technology inside these carts actually works, the main security vulnerabilities they face, and the privacy concerns they raise.

What's Inside the Cart?

A smart cart is essentially an IoT system made up of four components.

Product identification: Some systems use passive UHF (Ultra-High Frequency) RFID tags (operating in the 860-960 MHz band, with no battery on the tag itself). The tag harvests energy from the radio field emitted by the reader on the cart and reflects the incoming signal back, modulated with its own EPC (Electronic Product Code) identity. This is called backscatter. Since dozens of tags may be read at once, the Q algorithm in the EPC Gen2 standard spreads the tags across random time slots to avoid collisions.

Other systems rely entirely on computer vision. Shopic's clip-on device feeds the camera feed into a model trained on a catalog of more than 50,000 SKUs (Stock Keeping Units), recognizing products in real time and largely able to tell apart items that are dropped in motion or stacked on top of each other.

Verification: Load cells in the base of the cart measure the resistance change in strain gauges within a Wheatstone bridge to calculate weight differences. This is compared against the registered weight of the scanned product to catch tricks like scanning a cheap item while placing an expensive one in the cart.

Connectivity: Carts typically combine two wireless technologies. Low-power BLE (Bluetooth Low Energy) works with in-store location beacons to detect which aisle the cart is in, while Wi-Fi carries heavier traffic, such as camera footage and payment data, to the store server.

Backend: The data the cart collects is first processed on an onboard edge computing unit, then sent to the store server and from there to a cloud connected to inventory and payment systems.

Each of these four layers is its own attack surface, and the security of the overall system is only as strong as the weakest link in that chain.

Security

Nearly all of the ten critical vulnerability categories OWASP defines for IoT devices apply directly to smart carts. The standouts:

RFID/NFC cloning: An attacker can clone a tag's identity so an expensive product appears to be a cheap (or free) one. The recommended defense is mutual authentication between tag and reader, along with a randomized session key that changes on every read. (Source: Springer)
Man-in-the-middle (MITM) attacks over the network: If the traffic between the cart and the server isn't secured with encryption, an attacker positioned in between can alter the data being sent, for example swapping the identity of a scanned item for a cheaper one before it reaches the store's server.
Sensor spoofing: Because a camera or a weight sensor alone can be fooled, serious systems cross-check multiple sensors against each other.

Camera-based surveillance: Cameras used in systems like Amazon's Just Walk Out don't only see what ends up in the cart; the privacy implications of that are covered in the next section.
Biometric data risk: In systems that use palm or facial recognition for payment, a breach of that data is a breach of information that, unlike a password, can never be changed. This isn't to suggest that a supermarket's data ends up in a law-enforcement database, rather it's a useful illustration of how unforgiving a leak of biometric data can be, and how large these databases can grow: the FBI's National Palm Print System currently holds more than 29 million records, and none of them can be "reset" the way a password can. (Source: SiliconANGLE)

Privacy

Security vulnerabilities are about what an attacker can do; privacy concerns show up even when the system is working exactly as intended. One expert has described smart carts as "not a convenience, but a genuine surveillance technology," since these systems can log not only what a customer buys, but items they pick up, inspect, and put back without purchasing. (Source: SiliconANGLE)

This also has a legal dimension. Under the GDPR, biometric data (palm prints, facial recognition, and similar) is classified as "special category" personal data, subject to stricter requirements than ordinary consent. Regulators expect companies to justify why a biometric method was chosen in the first place, and to offer customers an alternative form of identification (a PIN or a card, for instance). In other words, "faster checkout" alone isn't treated as sufficient justification to require a less privacy-preserving method. In practice, this comes down to two principles: data minimization, meaning you don't collect data you don't need, or you delete it as soon as possible; and purpose limitation, meaning camera footage collected for "product recognition" shouldn't quietly get reused for customer profiling or ad targeting.

The Vendor Landscape

The different approaches in the industry also paint an interesting picture: Amazon Dash Cart uses a camera plus weight combination, Caper AI adds personalized ads on top of manual scanning, Veeve evolved over time into a retail media platform (Source: CNBC), and Shopic is entirely camera-based, eliminating scanning altogether; the company reported that smart cart usage in its live deployments increased customer spend by roughly 8%. (Source: Forbes, Source: Supermarket News) Each approach has its own attack surface: camera-heavy systems raise more privacy concerns, while scan-based systems are more exposed to tag fraud.

So What Should Consumers or Developers Do?

As a consumer, there are a few practical things worth checking before you rely on one of these carts:

Look for the store's stated retention policy: how long is your shopping and camera data kept, and is it shared with third parties (like advertisers)?
If a store offers palm or face-based payment, check whether a non-biometric option (PIN or card) is still available, and default to that if you're not comfortable with biometric enrollment.
Treat "convenience" features, like personalized in-cart ads, as a sign that your shopping data is also being used for purposes beyond the checkout itself.

For a developer or a security team, there are some concrete takeaways:

Never trust data coming from the client. Critical information like price must always be verified server-side, from your own source of truth.
Every message on the network needs integrity, and ideally confidentiality (TLS, i.e. Transport Layer Security, plus mutual authentication).
Critical decisions shouldn't rely on a single sensor; they should be based on cross-validation across multiple sensors.
Sensitive data such as camera footage or location should be collected under the principle of data minimization and with explicit consent. If biometric payment is offered, an alternative method (PIN or card) must also be available.

Closing

Smart carts genuinely improve the shopping experience, but being "smart" doesn't make them automatically secure. If anything, it's the opposite: every new sensor and every new network connection is one more attack surface. As a consumer, the key question is simple: what data is being collected in exchange for this convenience, and who holds onto it, and for how long? As a developer, the point is even more direct: security isn't a layer you bolt on afterward, it has to be part of the design from day one.

DEV Community