mitigate the risk in this scenario is to sandbox the model execution.

### Hardware attacks

Physical GPUs or TPUs can also be the target of attacks. [Published
research](https://scholar.google.com/scholar?q=gpu+side+channel) shows that it
might be possible to use side channel attacks on the GPU to leak data from other
running models or processes in the same system. GPUs can also have

Uit enquêtes onder Python ontwikkelaars blijkt maar al te duidelijk dat "(automatische) code aanvulling" <a href="https://www.jetbrains.com/research/python-developers-survey-2017/#tools-and-features" class="external-link" target="_blank">een van de meest gebruikte functionaliteiten is</a>.

 * hook into this and this is the model we need to follow - how many of our REST services do not map
 * directory to a method in the application interface?
 *
 * - research extension points versus plugins
 *
 * from igor: two plugins A and B, both depend on the same library but use different versions, say

			// Do not need to send any internal metadata
			// values to client.
			toRemove = append(toRemove, k)
			continue
		}

		// https://github.com/google/security-research/security/advisories/GHSA-76wf-9vgp-pj7w
		if equals(k, xhttp.AmzMetaUnencryptedContentLength, xhttp.AmzMetaUnencryptedContentMD5) {
			toRemove = append(toRemove, k)
			continue
		}
	}

	// whole checksums for many years now, no need to allocate.

	fi.Metadata = make(map[string]string, len(j.MetaUser)+len(j.MetaSys))
	for k, v := range j.MetaUser {
		// https://github.com/google/security-research/security/advisories/GHSA-76wf-9vgp-pj7w
		if equals(k, xhttp.AmzMetaUnencryptedContentLength, xhttp.AmzMetaUnencryptedContentMD5) {
			continue
		}
		if equals(k, "x-amz-storage-class") && v == storageclass.STANDARD {

							<c:if test="${s.index < 3}">
								<la:link
									href="/search?q=${f:u(item)}${fe:facetQuery()}${fe:geoQuery()}">${f:h(item)}</la:link>
							</c:if>
							<c:if test="${3 <= s.index}">
								<la:link styleClass="d-none d-sm-inline"
									href="/search?q=${f:u(item)}${fe:facetQuery()}${fe:geoQuery()}">${f:h(item)}</la:link>
							</c:if>
						</c:forEach>
					</p>

							<c:if test="${s.index < 3}">
								<la:link
									href="/search?q=${f:u(item)}${fe:facetQuery()}${fe:geoQuery()}">${f:h(item)}</la:link>
							</c:if>
							<c:if test="${3 <= s.index}">
								<la:link styleClass="d-none d-sm-inline"
									href="/search?q=${f:u(item)}${fe:facetQuery()}${fe:geoQuery()}">${f:h(item)}</la:link>
							</c:if>
						</c:forEach>
					</p>

		if stringsHasPrefixFold(k, ReservedMetadataPrefixLower) {
			continue
		}

		if equals(k, xhttp.AmzBucketReplicationStatus) {
			continue
		}

		// https://github.com/google/security-research/security/advisories/GHSA-76wf-9vgp-pj7w
		if equals(k, xhttp.AmzMetaUnencryptedContentLength, xhttp.AmzMetaUnencryptedContentMD5) {
			continue
		}
		meta[k] = v
	}

	if oi.ContentEncoding != "" {

*   Adds `beta` hyperparameter to
    [FTRL](https://www.tensorflow.org/api_docs/python/tf/keras/optimizers/Ftrl)
    optimizer classes (Keras and others) to match
    [FTRL paper](https://research.google.com/pubs/archive/41159.pdf).
*   `Optimizer.__init__` now accepts a `gradient_aggregator` to allow for
    customization of how gradients are aggregated across devices, as well as

authentication.admin.users=admin
authentication.admin.roles=admin

role.search.default.permissions=
role.search.default.display.permissions={role}guest
role.search.guest.permissions={role}guest

role.search.user.prefix=1
role.search.group.prefix=2
role.search.role.prefix=R
role.search.denied.prefix=D

# ----------------------------------------------------------