Antispoofing detection

Anti-spoofing detection model is a machine learning model designed to identify and prevent spoofing attacks by distinguishing between genuine and fake or manipulated inputs. In biometric systems, such as facial recognition or fingerprint scanning, spoofing involves presenting counterfeit data - like photos, masks, or synthetic fingerprints—to trick the system into granting unauthorized access. An anti-spoofing detection model analyzes subtle patterns, textures, movements, or other cues to detect whether the input originates from a live, authentic source or a spoofed attempt. This type of model is critical for enhancing the security and reliability of authentication systems.

At NeuralSpike we designed an anti-spotting model for systems that use face recognition. Our model is targeted for embedded devices like System On Module (SOM) or even Microcontroller Units (MCU). Our solution is based on MobileNetV3, with custom changes to the model architecture.

Our solution was deployed to GreenWaves GAP9 MCU. The GAP9 processor is great for embedded applications, and its main feature is low power consumption. In such a scenario our model was required to use less than 1MB of the memory, which includes model parameters and intermediate inference data. We used our ml-toolbox to find architecture that will work efficiently on GAP9, and that is able to extract maximum efficiency from its Neural Processing Accelerator - NE16. Our pipeline allowed us to find optimal architecture, and quantize it into integers.

To keep the power consumption at check, the GAP9 was connected to a camera that offered low resolution, and unique image characteristics. Thus we had to collect a train/test dataset to answer the above challenges. The data collection is very often a bottle neck of ML pipeline. First the data has to be collected with different people and surrounding environments, to cover the data variability. On top of that it is often very difficult to collect spoofing events, as they are rare. To address the above challenges we took advantage of our synthetic data generation framework. Our framework allows us to modify the input image with the prompt. We can change the appearance of the person, surrounding environment, and many more. In addition we can control what is changed in the image, making sure that the annotations are still valid. This way the generated images do not have to be annotated again. This is a huge win because manual annotations are a bottle neck for the whole process, and generate additional cost.

Thanks to our synthetic generation framework we needed to collect and annotate only a couple of hundred images. And later we multiplied the size of the training generating more images with our framework.

We also deployed our anti-spoofing model to the MediaTek Genio 700 platform. The Genio platform is much more powerful than GAP9 in terms of compute. Thus in this case we re-launch our architecture search, and found another architecture that fit better to the Genio hardware. This case is a good example which shows that our model training framework allows us to find optimal solutions for different hardware platforms.

Request your AI consultation

Name*

Last name*

Email*

Message*

Consent*

I consent to the processing of my personal data for the purpose of contact in connection with the inquiry or request submitted via the contact form. Consent to the processing of personal data is voluntary but necessary to receive a response. This consent may be withdrawn at any time without affecting the lawfulness of processing carried out before its withdrawal. The administrator of personal data is Neural Spike.