MindSpore Builds Huawei Cloud OCR Service with the Strict Security Assessment
MindSpore Builds Huawei Cloud OCR Service with the Strict Security Assessment
Cloud computing has become a new service and delivery mode in IT sector, but existing techniques have changed to fit new technologies, and this has created greater concern for security. This has led to government, industry associations, and other organizations to develop new policies, laws, regulations, and standards for cloud security. These compliance criteria aim to tighten security control over cloud computing to gain customer confidence and promote healthy and structural industry development.
Germany's Federal Office for Information Security (BSI), the national cybersecurity authority, developed the standard Cloud Computing Compliance Controls Catalogue (C5) to define basic security of cloud computing. C5 is considered the most comprehensive and strict data protection standard in the cloud industry, covering the information security measures to be met by cloud service providers. As AI and cloud computing are more integrated, AI technologies are posing new challenges to enterprises and individuals alike. AI features like generality, algorithm black box, and data dependency, are complex and involve huge volumes of data, and it is becoming clear the C5 standard alone is insufficient.
In 2021, the BSI, introduced the AI Cloud Service Compliance Criteria Catalogue (AIC4) for AI cloud services. This criteria, the first of its kind, defines a basic level of security for independent testing of the security of AI-based services throughout the lifecycle. In December, with the support of MindSpore, Huawei Cloud Optical Character Recognition (OCR) service successfully passed the AIC4 evaluation, verifying compliance and ability to provide secure cloud services.
OCR uses optical devices to capture images and automatically recognize characters, reducing labor costs and improving efficiency. However, not all OCR models are bulletproof. If inputs with small perturbations (adversarial examples) occur, some OCR models cannot correctly recognize or process character modifications, causing the text recognition accuracy to decrease, and making it difficult to locate the fault because the perturbations are imperceptible to naked eyes.
MindSpore provides the following built-in technical capabilities needed for lifecycle security:
Model security: protects models with robustness evaluation, adversarial evaluation and training, and model encryption.
Privacy protection: develops privacy technologies such as assessment, differential training, and federated learning.
Interpretability: improves the interpretability of models with TB-Net, the innovative semantic-level interpretable technology and explainable method toolset. Users can better understand, trust, and use models.
Huawei's technical team used MindSpore Armour, a model security evaluation tool, to evaluate the OCR security. Results showed algorithm models of most OCR services offer poor identification of natural perturbation and defense against adversarial examples. For example, the text box detection model has just 66% accuracy when attacked by noise check, PGD, and PSO.
The team has developed technologies, such as adversarial example detection and data augmentation training, which offer over 95% accuracy to identify adversarial examples. Huawei technologies improve model and OCR services and meet the security, robustness, and other criteria of AIC4.
MindSpore is not only key to ensuring Huawei Cloud OCR service could pass the AIC4 certification, but has also worked with authoritative audit organizations in France and Netherlands to develop the evaluation standards for AI frameworks. In April 2022, MindSpore was accredited with the CC EAL2+ certificate issued by the Netherlands Scheme for Certification in the Area of IT Security (NSCIB).
In the intelligent era, MindSpore is taking trustworthy AI to new heights. By continuously enhancing AI security in terms of technologies and governance, MindSpore is shaping the future together with partners, and building trustworthy AI that addresses AI security challenges.
For details about the AIC4, visit