MikeDoes (Michael Anthony)

posted an update 4 minutes ago

Post

The tools we use to audit AI for privacy might be easier to fool than we think.

We're highlighting a critical paper that introduces "PoisonM," a novel attack that could make Membership Inference tests unreliable. The direct connection to our work is explicit: the researchers, Neal M., Atul Prakash, Amrita Roy Chowdhury, Ashish Hooda, Kassem Fawaz, Somesh Jha, Zhuohang Li, and Brad Malin used the AI4Privacy dataset as the "canary" dataset in their experiments to test the effectiveness of their attack on realistic, sensitive information.

This is the power of a healthy open-source ecosystem. We provide the foundational data that helps researchers pressure-test our collective assumptions about AI safety. It's a win for everyone when this leads to a more honest conversation about what our tools can and can't do, pushing us all to create better solutions.

🔗 Read the full paper to understand the fundamental flaws in current MI testing: https://arxiv.org/pdf/2506.06003

#OpenSource
#DataPrivacy
#LLM
#Anonymization
#AIsecurity
#HuggingFace
#Ai4Privacy
#Worldslargestopensourceprivacymaskingdataset

reacted to their post with 👀 3 days ago

Post

3905

What if an AI agent could be tricked into stealing your data, just by reading a tool's description? A new paper reports it's possible.

The "Attractive Metadata Attack" paper details this stealthy new threat. To measure the real-world impact of their attack, the researchers needed a source of sensitive data for the agent to leak. We're proud that the AI4Privacy corpus was used to create the synthetic user profiles containing standardized PII for their experiments.

This is a perfect win-win. Our open-source data helped researchers Kanghua Mo, 龙昱丞, Zhihao Li from Guangzhou University and The Hong Kong Polytechnic University to not just demonstrate a new attack, but also quantify its potential for harm. This data-driven evidence is what pushes the community to build better, execution-level defenses for AI agents.

🔗 Check out their paper to see how easily an agent's trust in tool metadata could be exploited: https://arxiv.org/pdf/2508.02110

#OpenSource
#DataPrivacy
#LLM
#Anonymization
#AIsecurity
#HuggingFace
#Ai4Privacy
#Worldslargestopensourceprivacymaskingdataset

posted an update 3 days ago

Post

3905

What if an AI agent could be tricked into stealing your data, just by reading a tool's description? A new paper reports it's possible.

The "Attractive Metadata Attack" paper details this stealthy new threat. To measure the real-world impact of their attack, the researchers needed a source of sensitive data for the agent to leak. We're proud that the AI4Privacy corpus was used to create the synthetic user profiles containing standardized PII for their experiments.

This is a perfect win-win. Our open-source data helped researchers Kanghua Mo, 龙昱丞, Zhihao Li from Guangzhou University and The Hong Kong Polytechnic University to not just demonstrate a new attack, but also quantify its potential for harm. This data-driven evidence is what pushes the community to build better, execution-level defenses for AI agents.

🔗 Check out their paper to see how easily an agent's trust in tool metadata could be exploited: https://arxiv.org/pdf/2508.02110

#OpenSource
#DataPrivacy
#LLM
#Anonymization
#AIsecurity
#HuggingFace
#Ai4Privacy
#Worldslargestopensourceprivacymaskingdataset

reacted to their post with 👀🚀 6 days ago

Post

1797

How do you protect your prompts without breaking them? You need a smart sanitizer. A new system called Prϵϵmpt shows how.

The first, critical step in their solution is a high-performance Named Entity Recognition (NER) model to find the sensitive data. We're proud to see that these researchers, Amrita Roy Chowdhury, David Glukhov, Divyam Anshumaan, Prasad Chalasani, Nicolas Papernot, Somesh Jha, and Mihir Bellare from the University of Michigan, University of Toronto, University of Wisconsin-Madison, University of California, San Diego - Rady School of Management and Langroid Incorporated fine-tuned their NER model on 10 high-risk categories from the AI4Privacy dataset.

This is a perfect win-win. Our open-source data helps provide the foundation for the critical detection engine, which in turn enables the community to build and test better solutions like Prϵϵmpt's innovative use of encryption and Differential Privacy.

🔗 Check out their paper for a deep dive into a formally private, high-utility prompt sanitizer: https://arxiv.org/pdf/2504.05147

#OpenSource
#DataPrivacy
#LLM
#Anonymization
#AIsecurity
#HuggingFace
#Ai4Privacy
#Worldslargestopensourceprivacymaskingdataset

posted an update 7 days ago

Post

1797

How do you protect your prompts without breaking them? You need a smart sanitizer. A new system called Prϵϵmpt shows how.

The first, critical step in their solution is a high-performance Named Entity Recognition (NER) model to find the sensitive data. We're proud to see that these researchers, Amrita Roy Chowdhury, David Glukhov, Divyam Anshumaan, Prasad Chalasani, Nicolas Papernot, Somesh Jha, and Mihir Bellare from the University of Michigan, University of Toronto, University of Wisconsin-Madison, University of California, San Diego - Rady School of Management and Langroid Incorporated fine-tuned their NER model on 10 high-risk categories from the AI4Privacy dataset.

This is a perfect win-win. Our open-source data helps provide the foundation for the critical detection engine, which in turn enables the community to build and test better solutions like Prϵϵmpt's innovative use of encryption and Differential Privacy.

🔗 Check out their paper for a deep dive into a formally private, high-utility prompt sanitizer: https://arxiv.org/pdf/2504.05147

#OpenSource
#DataPrivacy
#LLM
#Anonymization
#AIsecurity
#HuggingFace
#Ai4Privacy
#Worldslargestopensourceprivacymaskingdataset

reacted to their post with 🔥 14 days ago

Post

3131

Making LLMs fast with KV-cache sharing is great. A new paper reports it's also a huge privacy risk.

That's why we're excited to see the "SafeKV" paper from researchers at the University of Connecticut, Peking University, and others. Their solution-oriented framework selectively shares non-sensitive data while isolating PII. To validate the "Safe" part of their system, they needed a robust, multilingual privacy benchmark.

We're proud that the Ai4Privacy pii-masking dataset was used for this critical evaluation related to privacy.

This is a perfect win-win. Our open-source data enables researchers to build and validate more effective security solutions for core AI infrastructure. Their work, in turn, helps make the entire LLM ecosystem safer, showing that performance and privacy don't have to be mutually exclusive.

Kudos to Kexin Chu, Zecheng Lin, Dawei Xiang, 沈子旭, Jianchang Su, cheng chu, Yiwei Yang, Wenhui Zhang, Wenfei Wu, and Wei Zhang on this beautiful work.

🔗 Check out their paper to see the future of secure, high-performance LLM inference: https://arxiv.org/pdf/2508.08438

#OpenSource
#DataPrivacy
#LLM
#Anonymization
#AIsecurity
#HuggingFace
#Ai4Privacy
#Worldslargestopensourceprivacymaskingdataset

posted an update 14 days ago

Post

3131

Making LLMs fast with KV-cache sharing is great. A new paper reports it's also a huge privacy risk.

That's why we're excited to see the "SafeKV" paper from researchers at the University of Connecticut, Peking University, and others. Their solution-oriented framework selectively shares non-sensitive data while isolating PII. To validate the "Safe" part of their system, they needed a robust, multilingual privacy benchmark.

We're proud that the Ai4Privacy pii-masking dataset was used for this critical evaluation related to privacy.

This is a perfect win-win. Our open-source data enables researchers to build and validate more effective security solutions for core AI infrastructure. Their work, in turn, helps make the entire LLM ecosystem safer, showing that performance and privacy don't have to be mutually exclusive.

Kudos to Kexin Chu, Zecheng Lin, Dawei Xiang, 沈子旭, Jianchang Su, cheng chu, Yiwei Yang, Wenhui Zhang, Wenfei Wu, and Wei Zhang on this beautiful work.

🔗 Check out their paper to see the future of secure, high-performance LLM inference: https://arxiv.org/pdf/2508.08438

#OpenSource
#DataPrivacy
#LLM
#Anonymization
#AIsecurity
#HuggingFace
#Ai4Privacy
#Worldslargestopensourceprivacymaskingdataset

reacted to their post with 🚀 24 days ago

Post

321

Are you sure the open-source LLM model you just downloaded is safe?

A recent paper on "Privacy Backdoors" reports a new vulnerability where pre-trained models can be poisoned before fine-tuning them. This is a serious challenge for everyone building on open-source AI.

Instead of just pointing out problems, we believe in finding better solutions. To understand this threat, the researchers needed to test their attack on realistic data structures. They needed a dataset that could effectively simulate a high-stakes privacy attack, and we're proud that our Ai4Privacy dataset was used to provide this crucial benchmark. The paper reports that for our complex dataset, the privacy leakage on a non-poisoned model was almost zero. After the backdoor attack, that number reportedly jumped to 87%.

Ai4Privacy dataset provided a realistic benchmark for their research. Our dataset, composed of synthetic identities, helped them demonstrate how a poisoned model could dramatically amplify privacy leakage.

This is why we champion open source: it enables the community to identify these issues and develop better, safer solutions together.

Kudos to the research team behind this study: Yuxin Wen, Leo Marchyok, Sanghyun Hong, Jonas Geiping, Tom Goldstein, and Nicholas Carlini, Oregon State University, University of Maryland, Google DeepMind, and ELLIS Institute Tubingen & MPI Intelligent Systems.

🔗 Read the research to understand this new challenge: https://arxiv.org/pdf/2404.01231

#DataPrivacy #AI #OpenSource #Anonymization #MachineLearning #Ai4Privacy #Worldslargestopensourceprivacydataset

posted an update 5 months ago

Post

321

Are you sure the open-source LLM model you just downloaded is safe?

A recent paper on "Privacy Backdoors" reports a new vulnerability where pre-trained models can be poisoned before fine-tuning them. This is a serious challenge for everyone building on open-source AI.

Instead of just pointing out problems, we believe in finding better solutions. To understand this threat, the researchers needed to test their attack on realistic data structures. They needed a dataset that could effectively simulate a high-stakes privacy attack, and we're proud that our Ai4Privacy dataset was used to provide this crucial benchmark. The paper reports that for our complex dataset, the privacy leakage on a non-poisoned model was almost zero. After the backdoor attack, that number reportedly jumped to 87%.

Ai4Privacy dataset provided a realistic benchmark for their research. Our dataset, composed of synthetic identities, helped them demonstrate how a poisoned model could dramatically amplify privacy leakage.

This is why we champion open source: it enables the community to identify these issues and develop better, safer solutions together.

Kudos to the research team behind this study: Yuxin Wen, Leo Marchyok, Sanghyun Hong, Jonas Geiping, Tom Goldstein, and Nicholas Carlini, Oregon State University, University of Maryland, Google DeepMind, and ELLIS Institute Tubingen & MPI Intelligent Systems.

🔗 Read the research to understand this new challenge: https://arxiv.org/pdf/2404.01231

#DataPrivacy #AI #OpenSource #Anonymization #MachineLearning #Ai4Privacy #Worldslargestopensourceprivacydataset

reacted to their post with 👀 5 months ago

Post

318

When anonymizing data for LLMs, is replacing a name with XXXXX enough?

A great post by Franklin Cardenoso Fernandez argues that we can do better. While simple masking hides data, it often destroys the context that models need to perform well.

A more robust method is contextual anonymization, where PII is replaced with meaningful labels like [NAME] or [ADDRESS]. This protects privacy while preserving the data's structural integrity.

We were pleased to see our Ai4Privacy pii-masking-200k dataset featured in the article as a prime example of this best practice. Our dataset is designed to help developers implement this superior form of anonymization by providing tens of thousands of clear, labeled examples.

By enabling models to be trained on data that is both private and context-rich, we can build AI that is both smarter and safer. This is a core part of our mission.

What's your team's preferred method for data anonymization? Let's discuss best practices.

🔗 Read Franklin's full analysis here: https://www.holisticai.com/blog/managing-personal-data-in-large-language-models

#DataPrivacy #Anonymization #ResponsibleAI #LLM #MachineLearning #AIEthics #Ai4Privacy #World's largest open privacy masking dataset

posted an update 5 months ago

Post

318

When anonymizing data for LLMs, is replacing a name with XXXXX enough?

A great post by Franklin Cardenoso Fernandez argues that we can do better. While simple masking hides data, it often destroys the context that models need to perform well.

A more robust method is contextual anonymization, where PII is replaced with meaningful labels like [NAME] or [ADDRESS]. This protects privacy while preserving the data's structural integrity.

We were pleased to see our Ai4Privacy pii-masking-200k dataset featured in the article as a prime example of this best practice. Our dataset is designed to help developers implement this superior form of anonymization by providing tens of thousands of clear, labeled examples.

By enabling models to be trained on data that is both private and context-rich, we can build AI that is both smarter and safer. This is a core part of our mission.

What's your team's preferred method for data anonymization? Let's discuss best practices.

🔗 Read Franklin's full analysis here: https://www.holisticai.com/blog/managing-personal-data-in-large-language-models

#DataPrivacy #Anonymization #ResponsibleAI #LLM #MachineLearning #AIEthics #Ai4Privacy #World's largest open privacy masking dataset

reacted to their post with 😎🚀 5 months ago

Post

2030

🛡️ At Ai4Privacy, our goal is to empower researchers to build a safer AI ecosystem. Today, we're highlighting crucial research that does just that by exposing a new vulnerability.

The paper "Forget to Flourish" details a new model poisoning technique. It's a reminder that as we fine-tune LLMs, our anonymization and privacy strategies must evolve to counter increasingly sophisticated threats.

We're proud that the Ai4Privacy dataset was instrumental in this study. It served two key purposes:

Provided a Realistic Testbed: It gave the researchers access to a diverse set of synthetic and realistic PII samples in a safe, controlled environment.

Enabled Impactful Benchmarking: It allowed them to measure the actual effectiveness of their data extraction attack, proving it could compromise specific, high-value information.

This work reinforces our belief that progress in AI security is a community effort. By providing robust tools for benchmarking, we can collectively identify weaknesses and build stronger, more resilient systems. A huge congratulations to the authors on this important contribution.

🔗 Read the full paper: https://arxiv.org/html/2408.17354v1

#OpenSource #DataPrivacy #LLM #Anonymization #AIsecurity #HuggingFace #Ai4Privacy #World's largest open privacy masking dataset

posted an update 5 months ago

Post

2030

🛡️ At Ai4Privacy, our goal is to empower researchers to build a safer AI ecosystem. Today, we're highlighting crucial research that does just that by exposing a new vulnerability.

The paper "Forget to Flourish" details a new model poisoning technique. It's a reminder that as we fine-tune LLMs, our anonymization and privacy strategies must evolve to counter increasingly sophisticated threats.

We're proud that the Ai4Privacy dataset was instrumental in this study. It served two key purposes:

Provided a Realistic Testbed: It gave the researchers access to a diverse set of synthetic and realistic PII samples in a safe, controlled environment.

Enabled Impactful Benchmarking: It allowed them to measure the actual effectiveness of their data extraction attack, proving it could compromise specific, high-value information.

This work reinforces our belief that progress in AI security is a community effort. By providing robust tools for benchmarking, we can collectively identify weaknesses and build stronger, more resilient systems. A huge congratulations to the authors on this important contribution.

🔗 Read the full paper: https://arxiv.org/html/2408.17354v1

#OpenSource #DataPrivacy #LLM #Anonymization #AIsecurity #HuggingFace #Ai4Privacy #World's largest open privacy masking dataset

reacted to their post with 🧠👀 5 months ago

Post

1153

In data privacy, 92% accuracy is not an A-grade. Privacy AI needs to be better.

That's the stark takeaway from a recent benchmark by Diego Mouriño

(Making Science), who put today's top PII detection methods to the test on call center transcripts using the Ai4Privacy dataset.

They pitted cutting-edge LLMs (like GPT-4 & Gemini) against traditional systems (like Cloud DLPs). The results show that our trust in these tools might be misplaced.

📊 The Hard Numbers:

Even top-tier LLMs peaked at a reported 92% accuracy, leaving a potential dangerous 8% gap where your customer's data can leak. They particularly struggled with basics like 'last names' and 'street addresses'.

The old guard? Traditional rule-based systems reportedly achieved a shocking 50% accuracy. A coin toss with your customers' privacy.

This tells us that for privacy tasks, off-the-shelf accuracy is a vanity metric. The real metric is the cost of a single failure—one leaked name, one exposed address.

While no tool is perfect, some are better than others. Diego’s full analysis breaks down which models offer the best cost-to-accuracy balance in this flawed landscape. It's a must-read for anyone serious about building trustworthy AI.

#DataPrivacy #AI #LLM #RiskManagement #MetricsThatMatter #InfoSec

Find the full post here:
https://www.makingscience.com/blog/protecting-customer-privacy-how-to-remove-pii-from-call-center-transcripts/

Dataset:
ai4privacy/pii-masking-400k

posted an update 5 months ago

Post

1153

In data privacy, 92% accuracy is not an A-grade. Privacy AI needs to be better.

That's the stark takeaway from a recent benchmark by Diego Mouriño

(Making Science), who put today's top PII detection methods to the test on call center transcripts using the Ai4Privacy dataset.

They pitted cutting-edge LLMs (like GPT-4 & Gemini) against traditional systems (like Cloud DLPs). The results show that our trust in these tools might be misplaced.

📊 The Hard Numbers:

Even top-tier LLMs peaked at a reported 92% accuracy, leaving a potential dangerous 8% gap where your customer's data can leak. They particularly struggled with basics like 'last names' and 'street addresses'.

The old guard? Traditional rule-based systems reportedly achieved a shocking 50% accuracy. A coin toss with your customers' privacy.

This tells us that for privacy tasks, off-the-shelf accuracy is a vanity metric. The real metric is the cost of a single failure—one leaked name, one exposed address.

While no tool is perfect, some are better than others. Diego’s full analysis breaks down which models offer the best cost-to-accuracy balance in this flawed landscape. It's a must-read for anyone serious about building trustworthy AI.

#DataPrivacy #AI #LLM #RiskManagement #MetricsThatMatter #InfoSec

Find the full post here:
https://www.makingscience.com/blog/protecting-customer-privacy-how-to-remove-pii-from-call-center-transcripts/

Dataset:
ai4privacy/pii-masking-400k

reacted to their post with ❤️👀 7 months ago

Post

2726

Started

aistatuscodes as a new project to create codes to understand AI performance better.

Going to be posting daily here and on instagram until we get to 100m downloads :)
https://www.instagram.com/MikeDoesDo/

Follow along the journey!

Michael Anthony PRO

AI & ML interests

Recent Activity

Organizations

Michael Anthony PRO

AI & ML interests

Recent Activity

Organizations

MikeDoes's activity