Noisy Aggregate Functions

Overview

Noisy aggregate functions are functions that provide random, noisy approximations of common aggregations like sum(), count(), and approx_distinct() as well as sketches like approx_set(). By injecting random noise into results, noisy aggregation functions make it more difficult to determine or confirm the exact data that was aggregated.

While many of these functions resemble differential privacy mechanisms, neither the values returned by these functions nor the query results that incorporate these functions are differentially private in general. See Limitations below for more details. Users who wish to support a strong privacy guarantee should discuss with a suitable technical expert first.

Counts, Sums, and Averages

noisy_count_gaussian(col, noise_scale[, random_seed]) -> bigint()

Counts the non-NULL values in col and then adds a normally distributed random double value with 0 mean and standard deviation of noise_scale to the true count. The noisy count is post-processed to be non-negative and rounded to bigint.

If provided, random_seed is used to seed the random number generator. Otherwise, noise is drawn from a secure random.

  1. SELECT noisy_count_gaussian(orderkey, 20.0) FROM tpch.tiny.lineitem; -- 60179 (1 row)
  2. SELECT noisy_count_gaussian(orderkey, 20.0) FROM tpch.tiny.lineitem WHERE false; -- NULL (1 row)

Note

Unlike count(), this function returns NULL when the (true) count of col is 0.

Distinct counting can be performed using noisy_count_gaussian(DISTINCT col, ...), or with noisy_approx_distinct_sfm(). Generally speaking, noisy_count_gaussian() returns more accurate results but at a larger computational cost.

noisy_count_if_gaussian(col, noise_scale[, random_seed]) -> bigint()

Counts the TRUE values in col and then adds a normally distributed random double value with 0 mean and standard deviation of noise_scale to the true count. The noisy count is post-processed to be non-negative and rounded to bigint.

If provided, random_seed is used to seed the random number generator. Otherwise, noise is drawn from a secure random.

  1. SELECT noisy_count_if_gaussian(orderkey > 10000, 20.0) FROM tpch.tiny.lineitem; -- 50180 (1 row)
  2. SELECT noisy_count_if_gaussian(orderkey > 10000, 20.0) FROM tpch.tiny.lineitem WHERE false; -- NULL (1 row)

Note

Unlike count_if(), this function returns NULL when the (true) count is 0.

noisy_sum_gaussian(col, noise_scale, lower, upper[, random_seed]) -> double()

Calculates the sum over the input values in col and then adds a normally distributed random double value with 0 mean and standard deviation of noise_scale. Each value is clipped to the range of [lower, upper] before adding to the sum.

If provided, random_seed is used to seed the random number generator. Otherwise, noise is drawn from a secure random.

noisy_sum_gaussian(col, noise_scale[, random_seed]) -> double()

Calculates the sum over the input values in col and then adds a normally distributed random double value with 0 mean and standard deviation of noise_scale.

If provided, random_seed is used to seed the random number generator. Otherwise, noise is drawn from a secure random.

noisy_avg_gaussian(col, noise_scale, lower, upper[, random_seed]) -> double()

Calculates the average (arithmetic mean) of all the input values in col and then adds a normally distributed random double value with 0 mean and standard deviation of noise_scale. Each value is clipped to the range of [lower, upper] before averaging.

If provided, random_seed is used to seed the random number generator. Otherwise, noise is drawn from a secure random.

noisy_avg_gaussian(col, noise_scale[, random_seed]) -> double()

Calculates the average (arithmetic mean) of all the input values in col and then adds a normally distributed random double value with 0 mean and standard deviation of noise_scale.

If provided, random_seed is used to seed the random number generator. Otherwise, noise is drawn from a secure random.

Approximate Distinct Counting/Sketching

Noisy approximate distinct counting and sketching (analogous to the deterministic HyperLogLog Functions) is supported via the Sketch-Flip-Merge (SFM) data sketch [Hehir2023].

noisy_approx_set_sfm(col, epsilon[, buckets[, precision]]) -> SfmSketch()

Returns an SFM sketch of the input values in col. This is analogous to the approx_set() function, which returns a (deterministic) HyperLogLog sketch.

  • col supports many types, similar to HyperLogLog.

  • epsilon (double) is a positive number that controls the level of noise in the sketch, as described in [Hehir2023]. Smaller values of epsilon correspond to noisier sketches.

  • buckets (int) defaults to 4096.

  • precision (int) defaults to 24.

Note

Unlike approx_set(), this function returns NULL when col is empty. If this behavior is undesirable, use coalesce() with noisy_empty_approx_set_sfm().

noisy_approx_set_sfm_from_index_and_zeros(col_index, col_zeros, epsilon, buckets[, precision]) -> SfmSketch()

Returns an SFM sketch of the input values in col_index and col_zeros.

This is similar to noisy_approx_set_sfm() except that function calculates a Murmur3Hash128.hash64() of col, and calculates the SFM PCSA bucket index and number of trailing zeros as described in [FlajoletMartin1985]. In this function, the caller must explicitly calculate the hash bucket index and zeros themselves and pass them as arguments col_index and col_zeros.

  • col_index (bigint) must be in the range 0..buckets-1.

  • col_zeros (bigint) must be in the range 0..64. If it exceeds precision, it is cropped to precision-1.

  • epsilon (double) is a positive number that controls the level of noise in the sketch, as described in [Hehir2023]. Smaller values of epsilon correspond to noisier sketches.

  • buckets (int) is the number of buckets in the SFM PCSA sketch as described in [Hehir2023].

  • precision (int) defaults to 24.

Note

Like noisy_approx_set_sfm(), this function returns NULL when col_index or col_zeros is NULL. If this behavior is undesirable, use coalesce() with noisy_empty_approx_set_sfm().

noisy_approx_distinct_sfm(col, epsilon[, buckets[, precision]]) -> bigint()

Equivalent to cardinality(noisy_approx_set_sfm(col, epsilon, buckets, precision)), this returns the approximate cardinality (distinct count) of the column col. This is analogous to the (deterministic) approx_distinct() function.

Note

Unlike approx_distinct(), this function returns NULL when col is empty.

noisy_empty_approx_set_sfm(epsilon[, buckets[, precision]]) -> SfmSketch()

Returns an SFM sketch with no items in it. This is analogous to the empty_approx_set() function, which returns an empty (deterministic) HyperLogLog sketch.

  • epsilon (double) is a positive number that controls the level of noise in the sketch, as described in [Hehir2023]. Smaller values of epsilon correspond to noisier sketches.

  • buckets (int) defaults to 4096.

  • precision (int) defaults to 24.

cardinality(SfmSketch) -> bigint()

Returns the estimated cardinality (distinct count) of an SfmSketch object.

merge(SfmSketch) -> SfmSketch()

An aggregator function that returns a merged SfmSketch of the set union of individual SfmSketch objects, similar to merge(HyperLogLog).

  1. SELECT year, cardinality(merge(sketch)) AS annual_distinct_count
  2. FROM monthly_sketches
  3. GROUP BY 1

merge_sfm(ARRAY[SfmSketch, …]) -> SfmSketch()

A scalar function that returns a merged SfmSketch of the set union of an array of SfmSketch objects, similar to merge_hll().

  1. SELECT cardinality(merge_sfm(ARRAY[
  2. noisy_approx_set_sfm(col_1, 5.0),
  3. noisy_approx_set_sfm(col_2, 5.0),
  4. noisy_approx_set_sfm(col_3, 5.0)
  5. ])) AS distinct_count_over_3_cols
  6. FROM my_table

Limitations

While these functions resemble differential privacy mechanisms, the values returned by these functions are not differentially private in general. There are several important limitations to keep in mind if using these functions for privacy-preserving purposes, including:

  • All noisy aggregate functions return NULL when aggregating empty sets. This means a NULL return value noiselessly indicates the absence of data.

  • GROUP BY clauses used in combination with noisy aggregation functions reveal non-noisy information: the presence or absence of a group noiselessly indicates the presence or absence of data. See, e.g., [Wilkins2024].

  • Functions relying on floating-point noise may be susceptible to inference attacks such as those identified in [Mironov2012] and [Casacuberta2022].


[Casacuberta2022]

Casacuberta, S., Shoemate, M., Vadhan, S., & Wagaman, C. (2022). Widespread Underestimation of Sensitivity in Differentially Private Libraries and How to Fix It. In Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security (pp. 471-484).

[Hehir2023] (1,2,3,4,5)

Hehir, J., Ting, D., & Cormode, G. (2023). Sketch-Flip-Merge: Mergeable Sketches for Private Distinct Counting. In Proceedings of the 40th International Conference on Machine Learning (Vol. 202).

[Mironov2012]

Mironov, I. (2012). On significance of the least significant bits for differential privacy. In Proceedings of the 2012 ACM Conference on Computer and Communications Security (pp. 650-661).

[Wilkins2024]

Wilkins, A., Kifer, D., Zhang, D., & Karrer, B. (2024). Exact Privacy Analysis of the Gaussian Sparse Histogram Mechanism. Journal of Privacy and Confidentiality, 14 (1).

[FlajoletMartin1985]

Flajolet, P, Martin, G. N. (1985). Probabilistic Counting Algorithms for Data Base Applications. In Journal of Computer and System Sciences, 31:182–209, 1985