Difference between revisions of "Murre"
Jump to navigation
Jump to search
m |
m |
||
| (One intermediate revision by the same user not shown) | |||
| Line 1: | Line 1: | ||
| − | Adroll's Mechanism for User Reports with Regulated Epsilon (Murre) is designed to enable machine learning based on binary inputs with noise added. The Murre proposal keeps the activity trail local to the web client, which thus requires client processing for any processing.<ref>https://github.com/AdRoll/privacy/blob/main/MURRE.md</ref> | + | Adroll's Mechanism for User Reports with Regulated [[Privacy Budget|Epsilon]] (Murre) is designed to enable machine learning based on binary inputs with noise added. The Murre proposal keeps the activity trail local to the web client, which thus requires client processing for any processing.<ref>https://github.com/AdRoll/privacy/blob/main/MURRE.md</ref> |
Murre makes the following assumptions: | Murre makes the following assumptions: | ||
* Machine learning algorithms in ad tech require granular data to function well. | * Machine learning algorithms in ad tech require granular data to function well. | ||
* Ad tech uses models of high dimension, with sparse vectors. | * Ad tech uses models of high dimension, with sparse vectors. | ||
| − | * Differential | + | * [[Differential Privacy|Differential Privacy]] is a framework for understanding and protecting user privacy. |
* The data required for effective advertising is not particularly sensitive. | * The data required for effective advertising is not particularly sensitive. | ||
* Computation and network usage will be limited on the user agent. | * Computation and network usage will be limited on the user agent. | ||
Latest revision as of 19:47, 13 February 2022
Adroll's Mechanism for User Reports with Regulated Epsilon (Murre) is designed to enable machine learning based on binary inputs with noise added. The Murre proposal keeps the activity trail local to the web client, which thus requires client processing for any processing.[1]
Murre makes the following assumptions:
- Machine learning algorithms in ad tech require granular data to function well.
- Ad tech uses models of high dimension, with sparse vectors.
- Differential Privacy is a framework for understanding and protecting user privacy.
- The data required for effective advertising is not particularly sensitive.
- Computation and network usage will be limited on the user agent.
Impact
By limiting inputs to binary values, continuous values (e.g., price) are not supported unless they are binned into categories, which will impair the modeling process.
By disclosing audience information, publishers and buyers are leaking their intellectual property associated with that web client (e.g., high value customer) to anyone with access to the web client.
Open Questions
- What is the time-delay in providing the data necessary for publishers to optimize their revenue?
- What is the time-delay in providing the data necessary for marketers to optimize their media spend?