Azure AI Content Safety

By Ashwin Venugopal -

 




How Does Azure AI Content Safety Work?

Azure AI Content Safety is designed to work with both text and images, as well as content generated by AI.

It can identify and moderate inappropriate material. The visual capabilities of Content Safety are driven by Microsoft's Florence foundation model, which has been trained on billions of pairs of text and images.

The analysis of text employs natural language processing methods to enhance the understanding of subtlety and context. Azure AI Content Safety supports multiple languages and is capable of recognizing harmful content in both short and long formats. It is currently available in English, German, Spanish, French, Portuguese, Italian, and Chinese.

Azure AI Content Safety features include:

Safeguarding Text Content

  • Moderate text scans text across four categories: violence, hate speech, sexual content, and self-harm. A severity level from 0 to 6 is returned for each category. This level helps to prioritize what needs immediate attention by people, and how urgently. You can also create a blocklist to scan for terms specific to your situation.

  • Prompt shields is a unified API to identify and block jailbreak attacks from inputs to LLMs. It includes both user input and documents. These attacks are prompts to LLMs that attempt to bypass the model's in-built safety features. User prompts are tested to ensure the input to the LLM is safe. Documents are tested to ensure they don't contain unsafe instructions embedded within the text.

  • Protected material detection checks AI-generated text for protected text such as recipes, copyrighted song lyrics, or other original material.

  • Groundedness detection protects against inaccurate responses in AI-generated text by LLMs. Public LLMs use data available at the time they were trained. However, data can be introduced after the original training of the model or be built on private data. A grounded response is one where the model’s output is based on the source information. An ungrounded response is one where the model's output varies from the source information. Groundedness detection includes a reasoning option in the API response. This adds a reasoning field that explains any ungroundedness detection. However, reasoning increases processing time and costs.

Safeguarding Image Content

  • Moderate images scans for inappropriate content across four categories: violence, self-harm, sexual, and hate. A severity level is returned: safe, low, or high. You then set a threshold level of low, medium, or high. The combination of the severity and threshold level determines whether the image is allowed or blocked for each category.

  • Moderate multimodal content scans both images and text, including text extracted from an image using Optical Character Recognition (OCR). Content is analyzed across four categories: violence, hate speech, sexual content, and self-harm.

Custom Safety Solutions

  • Custom categories enables you to create your own categories by providing positive and negative examples, and training the model. Content can then be scanned according to your own category definitions.

  • Safety system message helps you to write effective prompts to guide an AI system's behavior.

Limitations

Azure AI Content Safety employs AI algorithms, which means it might not consistently identify inappropriate language. Occasionally, it may also block acceptable language due to its reliance on algorithms and machine learning for identifying problematic language.

It is essential to test and assess Azure AI Content Safety using real data prior to deployment. After deployment, it is crucial to keep monitoring the system to evaluate its performance accuracy.

Conclusion

We have successfully learnt about AI Content Safety features.







Comments

Post a Comment

Popular posts from this blog

Information Protection Scanner: Resolve Issues with Information Protection Scanner Deployment

By Ashwin Venugopal -
Image
  About If you are experiencing problems with the Microsoft Purview Information Protection scanner, check the health of your deployment by utilizing the Start-ScannerDiagnosticsPowerShell cmdlet to start the scanner diagnostic tool. The diagnostics tool checks the following details and then creates a log files with the results: Whether the database is up to date. Whether network URLs are accessible. Whether there's a valid authentication token and policy can be acquired. Whether the profile is defined in the Azure portal. Whether offline/online configuration exists and can be acquired. Whether the rules configured are valid.  The Start-ScannerDiagnostics command does not perform a comprehensive check of prerequisites. If you encounter problems with the scanner, you need to verify that your system meets the scanner's requirements, and that your scanner's configuration and installation are finished.  Verify Scanning Details Per Scanner Node and Repository Run the Get-ScanSt...
Read more

How AMI Store & Restore Works?

By Ashwin Venugopal -
Image
  CreateStoreImageTask The CreateStoreImageTask API stores an AMI as a single object in an S3 bucket. The API initiates a task that collects all the data from the AMI and its associated snapshots, subsequently using an S3 multipart upload to store this data in an S3 object. It gathers all elements of the AMI, including most of the non-Region-specific metadata and all EBS snapshots included in the AMI, and consolidates them into a single object in S3. During the upload process, the data is compressed to minimize the space utilized in S3, which means that the S3 object could be smaller than the total sizes of the snapshots in the AMI. If the account making this API call has access to AMI and snapshot tags, they will be retained. The S3 object shares the same ID as the AMI, except it has a .bin extension. Additionally, the S3 metadata tags for the object include the AMI name, AMI description, AMI registration date, AMI owner account, and a timestamp indicating when the store operation...
Read more

Create A Store Image Task

By Ashwin Venugopal -
Image
  Introduction When you save an AMI in an S3 bucket, a task for storing the image is initiated. This store image task allows you to track the progress and result of the procedure. Securing your AMIs It is crucial to verify that the S3 bucket is set up with adequate security to protect the content of the AMI and that this security is upheld for the entire duration that the AMI objects are stored in the bucket. If this cannot be achieved, it is advisable to refrain from using these APIs. Make sure that public access to the S3 bucket is prohibited. It is suggested to activate Server-side encryption for the S3 buckets where you keep the AMIs, even though it’s not mandatory. When AMI snapshots are transferred to the S3 object, they are sent over TLS connections. It is possible to save AMIs that contain encrypted snapshots; however, the snapshots will be decrypted during the storing process. Identify the source AMI used to create a new Amazon EC2 AMI You can determine the AMI that was us...
Read more