Qualcomm Patents a Way to Run AI Models Without Leaking Their Secrets

By Patentlyze Team · Updated Jul 24, 2026

Every time a chip runs an AI model, it gives off tiny physical signals, power fluctuations, timing patterns, electromagnetic pulses, that a clever attacker can use to reverse-engineer what the model is doing. Qualcomm's new patent tries to block that.

How Qualcomm's 'masked gadgets' guard your on-device AI

Imagine a safe that, every time you open it, makes a distinctive sound. Even if nobody sees inside, someone listening carefully could learn your combination over time. AI chips have a similar problem: when they process data, they leak information through physical side-effects like power usage and timing, and attackers can exploit these to steal the AI model itself or the data it's handling.

Qualcomm's patent describes a system of what it calls 'masked gadgets,' isolated execution containers that each run a separate AI model at the same time. The idea is that running multiple models concurrently in their own shielded compartments makes it much harder for an eavesdropper to figure out what any single model is doing, because the signals get mixed up and obscured.

This kind of protection matters most for AI running directly on a device, like your phone or a car's onboard computer, where physical access to the chip is possible and cloud-based security measures don't apply.

How each masked gadget isolates a separate AI model's execution

The patent addresses side-channel attacks (a class of exploit that doesn't break the software directly but instead spies on physical signals the hardware emits while computing, such as power consumption or timing variations) against machine learning models running on a chip.

The core mechanism is the masked gadget: a dedicated execution unit that wraps an ML model in a layer designed to obscure those physical signals. The processor allocates one masked gadget per model, then runs multiple models through their respective gadgets at the same time (concurrently).

Running them in parallel is key. When several masked gadgets are active simultaneously, the physical signals they produce overlap, making it far harder for an attacker to isolate the signature of any one model. Think of it like trying to pick out a single conversation in a loud restaurant.

• Allocation: the processor assigns a dedicated masked gadget to each ML model before execution begins.
• Concurrent execution: all assigned gadgets run together, not sequentially.
• Signal masking: the gadget layer obscures the power and timing patterns the chip emits during that model's computation.

What this means for AI security in phones and cars

On-device AI is growing fast, in phones, cars, and edge hardware, and those chips are physically accessible in ways that cloud servers are not. A patient attacker with the right equipment can monitor a chip's power draw over thousands of inferences and reconstruct a proprietary model or extract sensitive data it processes. That's not a theoretical risk; side-channel attacks on cryptographic chips have been demonstrated in the real world for decades, and AI chips face the same exposure.

For Qualcomm specifically, whose Snapdragon chips power a huge share of Android phones and are being used in automotive AI systems, building this protection at the hardware-allocation level (rather than bolting it on in software) could become a meaningful differentiator as on-device AI handles more sensitive tasks like health data, authentication, and autonomous driving decisions.

Editorial commentary on a publicly published patent application. Not legal advice.