Version 1.2.0

Bug fix for UEClient _perturb() function to guarantee $\varepsilon$-LDP. See #11 for more information.

If you are using UEClient/UEServer, you should upgrade to this version so that they properly guarantee $\varepsilon$-LDP.

Version 1.1.2

Fixed and updated package dependencies on pip

Version 1.1.1

Fixed a pip installation issue

Version 1.1.0

Added new frequency oracles:

• Implemented Google's RAPPOR under pure_ldp.frequency_oracles.rappor
• Hadamard Mechanism (HM) under pure_ldp.frequency_oracles.hadamard_mechanism
• Direct Encoding (DE) under pure_ldp.frequency_oracles.direct_encoding
• A heuristic version of OLH called Fast Local Hashing (FLH) under ```pure_ldp.frequency_oracles.local_hashing``
• Private Count Sketch (PCS) an older version of private sketching under pure_ldp.frequency_oracles.priv_count_sketch
• SketchRespones (SR) a sketching oracle that uses any generic frequency oracle to perturb sketch rows under pure_ldp.frequency_oracles.sketch_response
• Zero oracle that just predicts 0 no matter what is aggregated (under pure_ldp.frequency_oracles.zero_oracle) - Mainly used for testing/experimenting

Added new heavy hitter implementations:

• Apple's Sequence Fragment Puzzle (SFP) under pure_ldp.heavy_hitters.apple_sfp
• Bassily et als TreeHistogram under pure_ldp.heavy_hitters.treehistogram

Other additions:

• Added pre-processing techniques to estimating over the whole domain for any frequency oracle (see the estimate_all of pure_ldp.core._freq_oracle_server )
• Added helpers for frequency oracles and heavy hitters for simulations (under pure_ldp.simulations.helpers)
• Added a long list of experiments and examples (under pure_ldp.simulations.paper_experiments.py)

Bug Fixes and Performance Improvements:

• Fixed wrong perturbation probability being used in Apple's HCMS
• Fixed a whole range of bugs where various parameters of frequency oracles wouldn't update properly
• Performance improvements for local_hashing and unary_encoding
• Fixed bugs in PEM

Version 1.0.6

Bug fixes:

• Hadamard Response now works in low privacy mode (epsilon < 1)
• Hadamard response should properly work via PyPI

New:

• Hadamard Response server has an additional normalisation parameter -
  • If normalization = 0 then no normalisation is applied (default)
  • If normalization = 1, the estimates are clipped to 0 and normalised
  • if normalization=2 then the estimates are projected onto the probability simplex

Version 1.0.5

Changes:

• Added Apple's CMS algorithm under pure_ldp_frequency_oracles.apple_cms
• Added helper methods to create instances of FOs/HHs in pure_ldp.core
• PEMServer aggregate method now takes a single tuple as an argument (i.e the output of PEMClient.privatise)
• HadamardResponseClient requires hash_funcs generated from HadamardResponseServer

Fixes:

• Updated some docstrings
• Fixed bug with HadamardResponse not permuting data properly
• Fixed typos in the README
• Fixed bugs in examples.py

Version 1.0.4

Breaking changes:

• Package structure has changed, all frequency oracles are now under pure_ldp.frequency_oracles i.e for local hashing it is now from pure_ldp.frequency_oracles.local_hashing import LHClient, LHServer
• Heavy hitters are also in their own package (just PEM at the moment)

Bugfixes:

• Fixed examples in example.py
• Added docstrings to PEM package
• Fixed raising errors in base class methods

Version 1.0.3

• Fixed bugs in update_params() for FreqOracleServer and added update_params() for FreqOracleClient
• Reworked core module structure
• Updated some documentation

Version 1.0.2

• Performance Improvements
• Added Prefix Extending Method (PEM) heavy hitter algorithm that uses OLH
• Added hadamard_response frequency oracle, this is a wrapper of hadamard_response
• Added base classes FreqOracleClient, FreqOracleServer in pure_ldp.core that all Frequency Oracles inherit from
• Improved shared aggregation/estimation logic
• Added more helper methods for frequency oracle servers

Version 1.0.1

Bug fixes:

• Fixed README typos
• Converted exceptions into warnings when using a small privacy budget/estimating on small amounts of data
• Fixed long description in setup.py