/**
 * @octomil/browser — Privacy filters
 *
 * Differential privacy (gradient clipping + noise injection) and
 * quantization for communication-efficient federated learning.
 */
import type { WeightMap } from "./types.js";
/**
 * Clip gradients by L2 norm. If the L2 norm of the flattened weight map
 * exceeds `maxNorm`, scale all values down proportionally.
 */
export declare function clipGradients(delta: WeightMap, maxNorm: number): WeightMap;
/**
 * Add calibrated Gaussian noise for (epsilon, delta)-differential privacy.
 *
 * Noise std = sensitivity * sqrt(2 * ln(1.25/deltaDP)) / epsilon
 *
 * @param delta     Weight deltas to perturb.
 * @param epsilon   Privacy budget.
 * @param sensitivity  L2 sensitivity (typically the clipping norm).
 * @param deltaDP   DP delta parameter.
 */
export declare function addGaussianNoise(delta: WeightMap, epsilon: number, sensitivity: number, deltaDP: number): WeightMap;
export interface QuantizedWeightMap {
    [key: string]: {
        data: Int8Array | Int16Array;
        scale: number;
        zeroPoint: number;
    };
}
/**
 * Quantize weights to reduced precision (8 or 16 bit).
 * Uses min-max symmetric quantization.
 */
export declare function quantize(delta: WeightMap, bits?: 8 | 16): QuantizedWeightMap;
/**
 * Dequantize back to Float32Array.
 */
export declare function dequantize(quantized: QuantizedWeightMap): WeightMap;
//# sourceMappingURL=privacy.d.ts.map